Questions about the V-C Iteration in Gelly

[Xingcan Cui]

Hi all,

I got some question about the vertex-centric iteration in Gelly.

a)  It seems the postSuperstep method is called before the superstep barrier (I got different aggregate values of the same superstep in this method). Is this a bug? Or the design is just like that?

b) There is not setHalt method  for vertices. When no message received, a vertex just quit the next iteration. Should I manually send messages (like heartbeat) to keep the vertices active?

c) I think we may need an initialization method in the ComputeFunction.

Any opinions? Thanks.

Best,

Xingcan

[Vasiliki Kalavri]

Hi Xingcan,

On 7 February 2017 at 10:10, Xingcan Cui <[hidden email]> wrote:

Hi all,

I got some question about the vertex-centric iteration in Gelly.

a)  It seems the postSuperstep method is called before the superstep barrier (I got different aggregate values of the same superstep in this method). Is this a bug? Or the design is just like that?

​The postSuperstep() method is called inside the close() method of a RichCoGroupFunction that wraps the ComputeFunction. The close() method It is called after the last call to the coGroup() after each iteration superstep.

The aggregate values are not guaranteed to be consistent during the same superstep when they are computed. To retrieve an aggregate value for superstep i, you should use the getPreviousIterationAggregate() method in superstep i+1.

 

b) There is not setHalt method for vertices. When no message received, a vertex just quit the next iteration. Should I manually send messages (like heartbeat) to keep the vertices active?

​That's because vertex halting is implicitly controlled by the underlying delta iterations of Flink. ​A vertex will remain active as long as it receives a message or it updates its value, otherwise it will become inactive. The documentation on Gelly iterations [1] and DataSet iterations [2] might be helpful.

 

c) I think we may need an initialization method in the ComputeFunction.

​There exists a preSuperstep() method for initialization. This one will be executed once per superstep before the compute function is invoked for every vertex. Would this work for you?

 

Any opinions? Thanks.

Best,

Xingcan

​I hope this helps,

-Vasia.​

​[1]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/libs/gelly/iterative_graph_processing.html#vertex-centric-iterations

[2]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/batch/iterations.html​

[Xingcan Cui]

Hi Vasia,

thanks for your reply. It helped a lot and I got some new ideas.

a) As you said, I did use the getPreviousIterationAggregate() method in preSuperstep() of the next superstep.

However, if the (only?) global (aggregate) results can not be guaranteed to be consistency,  what should we

do with the postSuperstep() method?

b) Though we can active vertices by update method or messages, IMO, it may be more proper for users 

themselves to decide when to halt a vertex's iteration. Considering a complex algorithm that contains different 

phases inside a vertex-centric iteration. Before moving to the next phase (that should be synchronized), 

there may be some vertices that already finished their work in current phase and they just wait for others.

Users may choose the finished vertices to idle until the next phase, but rather than to halt them. 

Can we consider adding the voteToHalt() method and some internal variables to the Vertex/Edge class 

(or just create an "advanced" version of them) to make the halting more controllable? 

c) Sorry that I didn't make it clear before. Here the initialization means a "global" one that executes once 

before the iteration. For example, users may want to initialize the vertices' values by their adjacent edges

before the iteration starts. Maybe we can add an extra coGroupFunction to the configuration parameters 

and apply it before the iteration?

What do you think?

(BTW, I started a PR on FLINK-1526(MST Lib&Example). Considering the complexity, the example is not 

provided.)

Really appreciate for all your help.

Best,

Xingcan

On Thu, Feb 9, 2017 at 5:36 PM, Vasiliki Kalavri <[hidden email]> wrote:

Hi Xingcan,

On 7 February 2017 at 10:10, Xingcan Cui <[hidden email]> wrote:

Hi all,

I got some question about the vertex-centric iteration in Gelly.

a)  It seems the postSuperstep method is called before the superstep barrier (I got different aggregate values of the same superstep in this method). Is this a bug? Or the design is just like that?

​The postSuperstep() method is called inside the close() method of a RichCoGroupFunction that wraps the ComputeFunction. The close() method It is called after the last call to the coGroup() after each iteration superstep.

The aggregate values are not guaranteed to be consistent during the same superstep when they are computed. To retrieve an aggregate value for superstep i, you should use the getPreviousIterationAggregate() method in superstep i+1.

 

b) There is not setHalt method for vertices. When no message received, a vertex just quit the next iteration. Should I manually send messages (like heartbeat) to keep the vertices active?

​That's because vertex halting is implicitly controlled by the underlying delta iterations of Flink. ​A vertex will remain active as long as it receives a message or it updates its value, otherwise it will become inactive. The documentation on Gelly iterations [1] and DataSet iterations [2] might be helpful.

 

c) I think we may need an initialization method in the ComputeFunction.

​There exists a preSuperstep() method for initialization. This one will be executed once per superstep before the compute function is invoked for every vertex. Would this work for you?

 

Any opinions? Thanks.

Best,

Xingcan

​I hope this helps,

-Vasia.​

​[1]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/libs/gelly/iterative_graph_processing.html#vertex-centric-iterations

[2]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/batch/iterations.html​

[Vasiliki Kalavri]

Hi Xingcan,

On 9 February 2017 at 18:16, Xingcan Cui <[hidden email]> wrote:

Hi Vasia,

thanks for your reply. It helped a lot and I got some new ideas.

a) As you said, I did use the getPreviousIterationAggregate() method in preSuperstep() of the next superstep.

However, if the (only?) global (aggregate) results can not be guaranteed to be consistency,  what should we

do with the postSuperstep() method?

​The postSuperstep() method is analogous to the close() method in a RichFunction, which is typically used for cleanup.​

 

b) Though we can active vertices by update method or messages, IMO, it may be more proper for users 

themselves to decide when to halt a vertex's iteration. Considering a complex algorithm that contains different 

phases inside a vertex-centric iteration. Before moving to the next phase (that should be synchronized), 

there may be some vertices that already finished their work in current phase and they just wait for others.

Users may choose the finished vertices to idle until the next phase, but rather than to halt them. 

Can we consider adding the voteToHalt() method and some internal variables to the Vertex/Edge class 

(or just create an "advanced" version of them) to make the halting more controllable? 

​I suppose adding a voteToHalt() method is possible, but I'm not sure I see how that would make halting more controllable. If a vertex hasn't changed value or hasn't received a message, it has no work to do in the next iteration, so why keep it active? If in a later superstep, a previously inactive vertex receives a message, it will become active again. ​Is this what you're looking for or am I missing something?

 

c) Sorry that I didn't make it clear before. Here the initialization means a "global" one that executes once 

before the iteration. For example, users may want to initialize the vertices' values by their adjacent edges

before the iteration starts. Maybe we can add an extra coGroupFunction to the configuration parameters 

and apply it before the iteration?

​You can initialize the graph by using any Gelly transformation methods before starting the iteration, e.g. mapVertices, mapEdges, reduceOnEdges, etc.

Btw, a vertex can iterate over its edges inside the ComputeFunction using the getEdges() method. Initializing the vertex values with neighboring edges might not be a good idea if you have vertices with high degrees.​

​Cheers,

-Vasia.​

 

What do you think?

(BTW, I started a PR on FLINK-1526(MST Lib&Example). Considering the complexity, the example is not 

provided.)

Really appreciate for all your help.

Best,

Xingcan

On Thu, Feb 9, 2017 at 5:36 PM, Vasiliki Kalavri <[hidden email]> wrote:

Hi Xingcan,

On 7 February 2017 at 10:10, Xingcan Cui <[hidden email]> wrote:

Hi all,

I got some question about the vertex-centric iteration in Gelly.

a)  It seems the postSuperstep method is called before the superstep barrier (I got different aggregate values of the same superstep in this method). Is this a bug? Or the design is just like that?

​The postSuperstep() method is called inside the close() method of a RichCoGroupFunction that wraps the ComputeFunction. The close() method It is called after the last call to the coGroup() after each iteration superstep.

The aggregate values are not guaranteed to be consistent during the same superstep when they are computed. To retrieve an aggregate value for superstep i, you should use the getPreviousIterationAggregate() method in superstep i+1.

 

b) There is not setHalt method for vertices. When no message received, a vertex just quit the next iteration. Should I manually send messages (like heartbeat) to keep the vertices active?

​That's because vertex halting is implicitly controlled by the underlying delta iterations of Flink. ​A vertex will remain active as long as it receives a message or it updates its value, otherwise it will become inactive. The documentation on Gelly iterations [1] and DataSet iterations [2] might be helpful.

 

c) I think we may need an initialization method in the ComputeFunction.

​There exists a preSuperstep() method for initialization. This one will be executed once per superstep before the compute function is invoked for every vertex. Would this work for you?

 

Any opinions? Thanks.

Best,

Xingcan

​I hope this helps,

-Vasia.​

​[1]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/libs/gelly/iterative_graph_processing.html#vertex-centric-iterations

[2]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/batch/iterations.html​

[Xingcan Cui]

Hi Vasia,

b) As I said, when some vertices finished their work in current phase, they have nothing to do (no value updates, no message received, just like slept) but to wait for other vertices that have not finished (the current phase) yet. After that in the next phase, all the vertices should go back to work again and if there are some vertices become inactive in last phase, it could be hard to reactive them again by message since we even don't know which vertices to send to. The only solution is to keep all vertices active, whether by updating vertices values in each super step or sending heartbeat messages to vertices themselves (which will bring a lot of extra work to the MessageCombiner). 

c) I know it's not elegant or even an awful idea to store the edge info into vertex values. However, we can not change edge values or maintain states (even a pick or unpick mark) in edges during a vertex-centric iteration. Then what can we do if an algorithm really need that?

Thanks for your patience.

Best,

Xingcan

On Fri, Feb 10, 2017 at 4:50 PM, Vasiliki Kalavri <[hidden email]> wrote:

Hi Xingcan,

On 9 February 2017 at 18:16, Xingcan Cui <[hidden email]> wrote:

Hi Vasia,

thanks for your reply. It helped a lot and I got some new ideas.

a) As you said, I did use the getPreviousIterationAggregate() method in preSuperstep() of the next superstep.

However, if the (only?) global (aggregate) results can not be guaranteed to be consistency,  what should we

do with the postSuperstep() method?

​The postSuperstep() method is analogous to the close() method in a RichFunction, which is typically used for cleanup.​

 

b) Though we can active vertices by update method or messages, IMO, it may be more proper for users 

themselves to decide when to halt a vertex's iteration. Considering a complex algorithm that contains different 

phases inside a vertex-centric iteration. Before moving to the next phase (that should be synchronized), 

there may be some vertices that already finished their work in current phase and they just wait for others.

Users may choose the finished vertices to idle until the next phase, but rather than to halt them. 

Can we consider adding the voteToHalt() method and some internal variables to the Vertex/Edge class 

(or just create an "advanced" version of them) to make the halting more controllable? 

​I suppose adding a voteToHalt() method is possible, but I'm not sure I see how that would make halting more controllable. If a vertex hasn't changed value or hasn't received a message, it has no work to do in the next iteration, so why keep it active? If in a later superstep, a previously inactive vertex receives a message, it will become active again. ​Is this what you're looking for or am I missing something?

 

c) Sorry that I didn't make it clear before. Here the initialization means a "global" one that executes once 

before the iteration. For example, users may want to initialize the vertices' values by their adjacent edges

before the iteration starts. Maybe we can add an extra coGroupFunction to the configuration parameters 

and apply it before the iteration?

​You can initialize the graph by using any Gelly transformation methods before starting the iteration, e.g. mapVertices, mapEdges, reduceOnEdges, etc.

Btw, a vertex can iterate over its edges inside the ComputeFunction using the getEdges() method. Initializing the vertex values with neighboring edges might not be a good idea if you have vertices with high degrees.​

​Cheers,

-Vasia.​

 

What do you think?

(BTW, I started a PR on FLINK-1526(MST Lib&Example). Considering the complexity, the example is not 

provided.)

Really appreciate for all your help.

Best,

Xingcan

On Thu, Feb 9, 2017 at 5:36 PM, Vasiliki Kalavri <[hidden email]> wrote:

Hi Xingcan,

On 7 February 2017 at 10:10, Xingcan Cui <[hidden email]> wrote:

Hi all,

I got some question about the vertex-centric iteration in Gelly.

a)  It seems the postSuperstep method is called before the superstep barrier (I got different aggregate values of the same superstep in this method). Is this a bug? Or the design is just like that?

​The postSuperstep() method is called inside the close() method of a RichCoGroupFunction that wraps the ComputeFunction. The close() method It is called after the last call to the coGroup() after each iteration superstep.

The aggregate values are not guaranteed to be consistent during the same superstep when they are computed. To retrieve an aggregate value for superstep i, you should use the getPreviousIterationAggregate() method in superstep i+1.

 

b) There is not setHalt method for vertices. When no message received, a vertex just quit the next iteration. Should I manually send messages (like heartbeat) to keep the vertices active?

​That's because vertex halting is implicitly controlled by the underlying delta iterations of Flink. ​A vertex will remain active as long as it receives a message or it updates its value, otherwise it will become inactive. The documentation on Gelly iterations [1] and DataSet iterations [2] might be helpful.

 

c) I think we may need an initialization method in the ComputeFunction.

​There exists a preSuperstep() method for initialization. This one will be executed once per superstep before the compute function is invoked for every vertex. Would this work for you?

 

Any opinions? Thanks.

Best,

Xingcan

​I hope this helps,

-Vasia.​

​[1]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/libs/gelly/iterative_graph_processing.html#vertex-centric-iterations

[2]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/batch/iterations.html​

[Greg Hogan]

Hi Xingcan,

FLINK-1885 looked into adding a bulk mode to Gelly's iterative models.

As an alternative you could implement your algorithm with Flink operators and a bulk iteration. Most of the Gelly library is written with native operators.

Greg

On Fri, Feb 10, 2017 at 5:02 AM, Xingcan Cui <[hidden email]> wrote:

Hi Vasia,

b) As I said, when some vertices finished their work in current phase, they have nothing to do (no value updates, no message received, just like slept) but to wait for other vertices that have not finished (the current phase) yet. After that in the next phase, all the vertices should go back to work again and if there are some vertices become inactive in last phase, it could be hard to reactive them again by message since we even don't know which vertices to send to. The only solution is to keep all vertices active, whether by updating vertices values in each super step or sending heartbeat messages to vertices themselves (which will bring a lot of extra work to the MessageCombiner). 

c) I know it's not elegant or even an awful idea to store the edge info into vertex values. However, we can not change edge values or maintain states (even a pick or unpick mark) in edges during a vertex-centric iteration. Then what can we do if an algorithm really need that?

Thanks for your patience.

Best,

Xingcan

On Fri, Feb 10, 2017 at 4:50 PM, Vasiliki Kalavri <[hidden email]> wrote:

Hi Xingcan,

On 9 February 2017 at 18:16, Xingcan Cui <[hidden email]> wrote:

Hi Vasia,

thanks for your reply. It helped a lot and I got some new ideas.

a) As you said, I did use the getPreviousIterationAggregate() method in preSuperstep() of the next superstep.

However, if the (only?) global (aggregate) results can not be guaranteed to be consistency,  what should we

do with the postSuperstep() method?

​The postSuperstep() method is analogous to the close() method in a RichFunction, which is typically used for cleanup.​

 

b) Though we can active vertices by update method or messages, IMO, it may be more proper for users 

themselves to decide when to halt a vertex's iteration. Considering a complex algorithm that contains different 

phases inside a vertex-centric iteration. Before moving to the next phase (that should be synchronized), 

there may be some vertices that already finished their work in current phase and they just wait for others.

Users may choose the finished vertices to idle until the next phase, but rather than to halt them. 

Can we consider adding the voteToHalt() method and some internal variables to the Vertex/Edge class 

(or just create an "advanced" version of them) to make the halting more controllable? 

​I suppose adding a voteToHalt() method is possible, but I'm not sure I see how that would make halting more controllable. If a vertex hasn't changed value or hasn't received a message, it has no work to do in the next iteration, so why keep it active? If in a later superstep, a previously inactive vertex receives a message, it will become active again. ​Is this what you're looking for or am I missing something?

 

c) Sorry that I didn't make it clear before. Here the initialization means a "global" one that executes once 

before the iteration. For example, users may want to initialize the vertices' values by their adjacent edges

before the iteration starts. Maybe we can add an extra coGroupFunction to the configuration parameters 

and apply it before the iteration?

​You can initialize the graph by using any Gelly transformation methods before starting the iteration, e.g. mapVertices, mapEdges, reduceOnEdges, etc.

Btw, a vertex can iterate over its edges inside the ComputeFunction using the getEdges() method. Initializing the vertex values with neighboring edges might not be a good idea if you have vertices with high degrees.​

​Cheers,

-Vasia.​

 

What do you think?

(BTW, I started a PR on FLINK-1526(MST Lib&Example). Considering the complexity, the example is not 

provided.)

Really appreciate for all your help.

Best,

Xingcan

On Thu, Feb 9, 2017 at 5:36 PM, Vasiliki Kalavri <[hidden email]> wrote:

Hi Xingcan,

On 7 February 2017 at 10:10, Xingcan Cui <[hidden email]> wrote:

Hi all,

I got some question about the vertex-centric iteration in Gelly.

a)  It seems the postSuperstep method is called before the superstep barrier (I got different aggregate values of the same superstep in this method). Is this a bug? Or the design is just like that?

​The postSuperstep() method is called inside the close() method of a RichCoGroupFunction that wraps the ComputeFunction. The close() method It is called after the last call to the coGroup() after each iteration superstep.

The aggregate values are not guaranteed to be consistent during the same superstep when they are computed. To retrieve an aggregate value for superstep i, you should use the getPreviousIterationAggregate() method in superstep i+1.

 

b) There is not setHalt method for vertices. When no message received, a vertex just quit the next iteration. Should I manually send messages (like heartbeat) to keep the vertices active?

​That's because vertex halting is implicitly controlled by the underlying delta iterations of Flink. ​A vertex will remain active as long as it receives a message or it updates its value, otherwise it will become inactive. The documentation on Gelly iterations [1] and DataSet iterations [2] might be helpful.

 

c) I think we may need an initialization method in the ComputeFunction.

​There exists a preSuperstep() method for initialization. This one will be executed once per superstep before the compute function is invoked for every vertex. Would this work for you?

 

Any opinions? Thanks.

Best,

Xingcan

​I hope this helps,

-Vasia.​

​[1]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/libs/gelly/iterative_graph_processing.html#vertex-centric-iterations

[2]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/batch/iterations.html​

[Xingcan Cui]

Hi Greg,

Thanks for your attention.

It takes me a little time to read the old PR on FLINK-1885. Though the VertexCentricIteration, as well as its related classes, has been refactored, I understand what Markus want to achieve.

I am not sure if using a bulk iteration instead of a delta one could eliminate the "out of memory" problem.  Except for that, I think the "auto update" has nothing to do with the bulk mode. Considering the compatible guarantee, here is my suggestions to improve gelly's iteration API:

1) Add an "autoHalt" flag to the ComputeFunction.

2) If the flag is set true (default), apply the current mechanism .

3) If the flag is set false, call out.collect() to update the vertex value whether the setNewVertexValue() method is called or not, unless the user explicitly call a (new added) voteToHalt() method in the ComputeFunction.

By adding these, users can decide when to halt a vertex themselves. What do you think? 

As for the "update edge values during vertex iterations" problem, I think it needs a redesign for the gelly framework (Maybe merge the vertices and edges into a single data set? Or just change the iterations'  implementation? I can't think it clearly now.), so that's it for now. Besides, I don't think there will be someone who really would love to write a graph algorithm with Flink native operators and that's why gelly is designed, isn't it?

Best,

Xingcan

On Fri, Feb 10, 2017 at 10:31 PM, Greg Hogan <[hidden email]> wrote:

Hi Xingcan,

FLINK-1885 looked into adding a bulk mode to Gelly's iterative models.

As an alternative you could implement your algorithm with Flink operators and a bulk iteration. Most of the Gelly library is written with native operators.

Greg

On Fri, Feb 10, 2017 at 5:02 AM, Xingcan Cui <[hidden email]> wrote:

Hi Vasia,

b) As I said, when some vertices finished their work in current phase, they have nothing to do (no value updates, no message received, just like slept) but to wait for other vertices that have not finished (the current phase) yet. After that in the next phase, all the vertices should go back to work again and if there are some vertices become inactive in last phase, it could be hard to reactive them again by message since we even don't know which vertices to send to. The only solution is to keep all vertices active, whether by updating vertices values in each super step or sending heartbeat messages to vertices themselves (which will bring a lot of extra work to the MessageCombiner). 

c) I know it's not elegant or even an awful idea to store the edge info into vertex values. However, we can not change edge values or maintain states (even a pick or unpick mark) in edges during a vertex-centric iteration. Then what can we do if an algorithm really need that?

Thanks for your patience.

Best,

Xingcan

On Fri, Feb 10, 2017 at 4:50 PM, Vasiliki Kalavri <[hidden email]> wrote:

Hi Xingcan,

On 9 February 2017 at 18:16, Xingcan Cui <[hidden email]> wrote:

Hi Vasia,

thanks for your reply. It helped a lot and I got some new ideas.

a) As you said, I did use the getPreviousIterationAggregate() method in preSuperstep() of the next superstep.

However, if the (only?) global (aggregate) results can not be guaranteed to be consistency,  what should we

do with the postSuperstep() method?

​The postSuperstep() method is analogous to the close() method in a RichFunction, which is typically used for cleanup.​

 

b) Though we can active vertices by update method or messages, IMO, it may be more proper for users 

themselves to decide when to halt a vertex's iteration. Considering a complex algorithm that contains different 

phases inside a vertex-centric iteration. Before moving to the next phase (that should be synchronized), 

there may be some vertices that already finished their work in current phase and they just wait for others.

Users may choose the finished vertices to idle until the next phase, but rather than to halt them. 

Can we consider adding the voteToHalt() method and some internal variables to the Vertex/Edge class 

(or just create an "advanced" version of them) to make the halting more controllable? 

​I suppose adding a voteToHalt() method is possible, but I'm not sure I see how that would make halting more controllable. If a vertex hasn't changed value or hasn't received a message, it has no work to do in the next iteration, so why keep it active? If in a later superstep, a previously inactive vertex receives a message, it will become active again. ​Is this what you're looking for or am I missing something?

 

c) Sorry that I didn't make it clear before. Here the initialization means a "global" one that executes once 

before the iteration. For example, users may want to initialize the vertices' values by their adjacent edges

before the iteration starts. Maybe we can add an extra coGroupFunction to the configuration parameters 

and apply it before the iteration?

​You can initialize the graph by using any Gelly transformation methods before starting the iteration, e.g. mapVertices, mapEdges, reduceOnEdges, etc.

Btw, a vertex can iterate over its edges inside the ComputeFunction using the getEdges() method. Initializing the vertex values with neighboring edges might not be a good idea if you have vertices with high degrees.​

​Cheers,

-Vasia.​

 

What do you think?

(BTW, I started a PR on FLINK-1526(MST Lib&Example). Considering the complexity, the example is not 

provided.)

Really appreciate for all your help.

Best,

Xingcan

On Thu, Feb 9, 2017 at 5:36 PM, Vasiliki Kalavri <[hidden email]> wrote:

Hi Xingcan,

On 7 February 2017 at 10:10, Xingcan Cui <[hidden email]> wrote:

Hi all,

I got some question about the vertex-centric iteration in Gelly.

a)  It seems the postSuperstep method is called before the superstep barrier (I got different aggregate values of the same superstep in this method). Is this a bug? Or the design is just like that?

​The postSuperstep() method is called inside the close() method of a RichCoGroupFunction that wraps the ComputeFunction. The close() method It is called after the last call to the coGroup() after each iteration superstep.

The aggregate values are not guaranteed to be consistent during the same superstep when they are computed. To retrieve an aggregate value for superstep i, you should use the getPreviousIterationAggregate() method in superstep i+1.

 

b) There is not setHalt method for vertices. When no message received, a vertex just quit the next iteration. Should I manually send messages (like heartbeat) to keep the vertices active?

​That's because vertex halting is implicitly controlled by the underlying delta iterations of Flink. ​A vertex will remain active as long as it receives a message or it updates its value, otherwise it will become inactive. The documentation on Gelly iterations [1] and DataSet iterations [2] might be helpful.

 

c) I think we may need an initialization method in the ComputeFunction.

​There exists a preSuperstep() method for initialization. This one will be executed once per superstep before the compute function is invoked for every vertex. Would this work for you?

 

Any opinions? Thanks.

Best,

Xingcan

​I hope this helps,

-Vasia.​

​[1]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/libs/gelly/iterative_graph_processing.html#vertex-centric-iterations

[2]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/batch/iterations.html​

[Xingcan Cui]

Hi Greg,

I also found that in VertexCentricIteration.java, the message set is taken as the workset while the vertex set is taken as the delta for solution set. By doing like that, the setNewVertex method will not actually active a vertex. In other words, if no message is generated (the workset is empty) the "pact.runtime.workset-empty-aggregator" will judge convergence of the delta iteration and then the iteration just terminates. Is this a bug?

Best,

Xingcan

On Mon, Feb 13, 2017 at 5:24 PM, Xingcan Cui <[hidden email]> wrote:

Hi Greg,

Thanks for your attention.

It takes me a little time to read the old PR on FLINK-1885. Though the VertexCentricIteration, as well as its related classes, has been refactored, I understand what Markus want to achieve.

I am not sure if using a bulk iteration instead of a delta one could eliminate the "out of memory" problem.  Except for that, I think the "auto update" has nothing to do with the bulk mode. Considering the compatible guarantee, here is my suggestions to improve gelly's iteration API:

1) Add an "autoHalt" flag to the ComputeFunction.

2) If the flag is set true (default), apply the current mechanism .

3) If the flag is set false, call out.collect() to update the vertex value whether the setNewVertexValue() method is called or not, unless the user explicitly call a (new added) voteToHalt() method in the ComputeFunction.

By adding these, users can decide when to halt a vertex themselves. What do you think? 

As for the "update edge values during vertex iterations" problem, I think it needs a redesign for the gelly framework (Maybe merge the vertices and edges into a single data set? Or just change the iterations'  implementation? I can't think it clearly now.), so that's it for now. Besides, I don't think there will be someone who really would love to write a graph algorithm with Flink native operators and that's why gelly is designed, isn't it?

Best,

Xingcan

On Fri, Feb 10, 2017 at 10:31 PM, Greg Hogan <[hidden email]> wrote:

Hi Xingcan,

FLINK-1885 looked into adding a bulk mode to Gelly's iterative models.

As an alternative you could implement your algorithm with Flink operators and a bulk iteration. Most of the Gelly library is written with native operators.

Greg

On Fri, Feb 10, 2017 at 5:02 AM, Xingcan Cui <[hidden email]> wrote:

Hi Vasia,

b) As I said, when some vertices finished their work in current phase, they have nothing to do (no value updates, no message received, just like slept) but to wait for other vertices that have not finished (the current phase) yet. After that in the next phase, all the vertices should go back to work again and if there are some vertices become inactive in last phase, it could be hard to reactive them again by message since we even don't know which vertices to send to. The only solution is to keep all vertices active, whether by updating vertices values in each super step or sending heartbeat messages to vertices themselves (which will bring a lot of extra work to the MessageCombiner). 

c) I know it's not elegant or even an awful idea to store the edge info into vertex values. However, we can not change edge values or maintain states (even a pick or unpick mark) in edges during a vertex-centric iteration. Then what can we do if an algorithm really need that?

Thanks for your patience.

Best,

Xingcan

On Fri, Feb 10, 2017 at 4:50 PM, Vasiliki Kalavri <[hidden email]> wrote:

Hi Xingcan,

On 9 February 2017 at 18:16, Xingcan Cui <[hidden email]> wrote:

Hi Vasia,

thanks for your reply. It helped a lot and I got some new ideas.

a) As you said, I did use the getPreviousIterationAggregate() method in preSuperstep() of the next superstep.

However, if the (only?) global (aggregate) results can not be guaranteed to be consistency,  what should we

do with the postSuperstep() method?

​The postSuperstep() method is analogous to the close() method in a RichFunction, which is typically used for cleanup.​

 

b) Though we can active vertices by update method or messages, IMO, it may be more proper for users 

themselves to decide when to halt a vertex's iteration. Considering a complex algorithm that contains different 

phases inside a vertex-centric iteration. Before moving to the next phase (that should be synchronized), 

there may be some vertices that already finished their work in current phase and they just wait for others.

Users may choose the finished vertices to idle until the next phase, but rather than to halt them. 

Can we consider adding the voteToHalt() method and some internal variables to the Vertex/Edge class 

(or just create an "advanced" version of them) to make the halting more controllable? 

​I suppose adding a voteToHalt() method is possible, but I'm not sure I see how that would make halting more controllable. If a vertex hasn't changed value or hasn't received a message, it has no work to do in the next iteration, so why keep it active? If in a later superstep, a previously inactive vertex receives a message, it will become active again. ​Is this what you're looking for or am I missing something?

 

c) Sorry that I didn't make it clear before. Here the initialization means a "global" one that executes once 

before the iteration. For example, users may want to initialize the vertices' values by their adjacent edges

before the iteration starts. Maybe we can add an extra coGroupFunction to the configuration parameters 

and apply it before the iteration?

​You can initialize the graph by using any Gelly transformation methods before starting the iteration, e.g. mapVertices, mapEdges, reduceOnEdges, etc.

Btw, a vertex can iterate over its edges inside the ComputeFunction using the getEdges() method. Initializing the vertex values with neighboring edges might not be a good idea if you have vertices with high degrees.​

​Cheers,

-Vasia.​

 

What do you think?

(BTW, I started a PR on FLINK-1526(MST Lib&Example). Considering the complexity, the example is not 

provided.)

Really appreciate for all your help.

Best,

Xingcan

On Thu, Feb 9, 2017 at 5:36 PM, Vasiliki Kalavri <[hidden email]> wrote:

Hi Xingcan,

On 7 February 2017 at 10:10, Xingcan Cui <[hidden email]> wrote:

Hi all,

I got some question about the vertex-centric iteration in Gelly.

a)  It seems the postSuperstep method is called before the superstep barrier (I got different aggregate values of the same superstep in this method). Is this a bug? Or the design is just like that?

​The postSuperstep() method is called inside the close() method of a RichCoGroupFunction that wraps the ComputeFunction. The close() method It is called after the last call to the coGroup() after each iteration superstep.

The aggregate values are not guaranteed to be consistent during the same superstep when they are computed. To retrieve an aggregate value for superstep i, you should use the getPreviousIterationAggregate() method in superstep i+1.

 

b) There is not setHalt method for vertices. When no message received, a vertex just quit the next iteration. Should I manually send messages (like heartbeat) to keep the vertices active?

​That's because vertex halting is implicitly controlled by the underlying delta iterations of Flink. ​A vertex will remain active as long as it receives a message or it updates its value, otherwise it will become inactive. The documentation on Gelly iterations [1] and DataSet iterations [2] might be helpful.

 

c) I think we may need an initialization method in the ComputeFunction.

​There exists a preSuperstep() method for initialization. This one will be executed once per superstep before the compute function is invoked for every vertex. Would this work for you?

 

Any opinions? Thanks.

Best,

Xingcan

​I hope this helps,

-Vasia.​

​[1]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/libs/gelly/iterative_graph_processing.html#vertex-centric-iterations

[2]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/batch/iterations.html​

[Vasiliki Kalavri]

Hi Xingcan,

that's my bad, I was thinking of scatter-gather iterations in my previous reply. You're right, in VertexCentricIteration a vertex is only active in the next superstep if it has received at least one message in the current superstep. Updating its value does not impact the activation. This is intentional in the vertex-centric model.

I agree that the current design of the iterative models is restrictive and doesn't allow for the expression of complex iterative algorithms that require updating edges or defining phases. We have discussed this before, e.g. in [1]. The outcome of that discussion was that we should use for-loop iterations for such cases, as the closed-loop iteration operators of Flink might not provide the necessary flexibility. As you will see in the thread though, that proposal didn't work out, as efficiently supporting for-loops in Flink is not possible right now.

-Vasia.

[1]: http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/DISCUSS-Gelly-iteration-abstractions-td3949.html

On 14 February 2017 at 08:10, Xingcan Cui <[hidden email]> wrote:

Hi Greg,

I also found that in VertexCentricIteration.java, the message set is taken as the workset while the vertex set is taken as the delta for solution set. By doing like that, the setNewVertex method will not actually active a vertex. In other words, if no message is generated (the workset is empty) the "pact.runtime.workset-empty-aggregator" will judge convergence of the delta iteration and then the iteration just terminates. Is this a bug?

Best,

Xingcan

On Mon, Feb 13, 2017 at 5:24 PM, Xingcan Cui <[hidden email]> wrote:

Hi Greg,

Thanks for your attention.

It takes me a little time to read the old PR on FLINK-1885. Though the VertexCentricIteration, as well as its related classes, has been refactored, I understand what Markus want to achieve.

I am not sure if using a bulk iteration instead of a delta one could eliminate the "out of memory" problem.  Except for that, I think the "auto update" has nothing to do with the bulk mode. Considering the compatible guarantee, here is my suggestions to improve gelly's iteration API:

1) Add an "autoHalt" flag to the ComputeFunction.

2) If the flag is set true (default), apply the current mechanism .

3) If the flag is set false, call out.collect() to update the vertex value whether the setNewVertexValue() method is called or not, unless the user explicitly call a (new added) voteToHalt() method in the ComputeFunction.

By adding these, users can decide when to halt a vertex themselves. What do you think? 

As for the "update edge values during vertex iterations" problem, I think it needs a redesign for the gelly framework (Maybe merge the vertices and edges into a single data set? Or just change the iterations'  implementation? I can't think it clearly now.), so that's it for now. Besides, I don't think there will be someone who really would love to write a graph algorithm with Flink native operators and that's why gelly is designed, isn't it?

Best,

Xingcan

On Fri, Feb 10, 2017 at 10:31 PM, Greg Hogan <[hidden email]> wrote:

Hi Xingcan,

FLINK-1885 looked into adding a bulk mode to Gelly's iterative models.

As an alternative you could implement your algorithm with Flink operators and a bulk iteration. Most of the Gelly library is written with native operators.

Greg

On Fri, Feb 10, 2017 at 5:02 AM, Xingcan Cui <[hidden email]> wrote:

Hi Vasia,

b) As I said, when some vertices finished their work in current phase, they have nothing to do (no value updates, no message received, just like slept) but to wait for other vertices that have not finished (the current phase) yet. After that in the next phase, all the vertices should go back to work again and if there are some vertices become inactive in last phase, it could be hard to reactive them again by message since we even don't know which vertices to send to. The only solution is to keep all vertices active, whether by updating vertices values in each super step or sending heartbeat messages to vertices themselves (which will bring a lot of extra work to the MessageCombiner). 

c) I know it's not elegant or even an awful idea to store the edge info into vertex values. However, we can not change edge values or maintain states (even a pick or unpick mark) in edges during a vertex-centric iteration. Then what can we do if an algorithm really need that?

Thanks for your patience.

Best,

Xingcan

On Fri, Feb 10, 2017 at 4:50 PM, Vasiliki Kalavri <[hidden email]> wrote:

Hi Xingcan,

On 9 February 2017 at 18:16, Xingcan Cui <[hidden email]> wrote:

Hi Vasia,

thanks for your reply. It helped a lot and I got some new ideas.

a) As you said, I did use the getPreviousIterationAggregate() method in preSuperstep() of the next superstep.

However, if the (only?) global (aggregate) results can not be guaranteed to be consistency,  what should we

do with the postSuperstep() method?

​The postSuperstep() method is analogous to the close() method in a RichFunction, which is typically used for cleanup.​

 

b) Though we can active vertices by update method or messages, IMO, it may be more proper for users 

themselves to decide when to halt a vertex's iteration. Considering a complex algorithm that contains different 

phases inside a vertex-centric iteration. Before moving to the next phase (that should be synchronized), 

there may be some vertices that already finished their work in current phase and they just wait for others.

Users may choose the finished vertices to idle until the next phase, but rather than to halt them. 

Can we consider adding the voteToHalt() method and some internal variables to the Vertex/Edge class 

(or just create an "advanced" version of them) to make the halting more controllable? 

​I suppose adding a voteToHalt() method is possible, but I'm not sure I see how that would make halting more controllable. If a vertex hasn't changed value or hasn't received a message, it has no work to do in the next iteration, so why keep it active? If in a later superstep, a previously inactive vertex receives a message, it will become active again. ​Is this what you're looking for or am I missing something?

 

c) Sorry that I didn't make it clear before. Here the initialization means a "global" one that executes once 

before the iteration. For example, users may want to initialize the vertices' values by their adjacent edges

before the iteration starts. Maybe we can add an extra coGroupFunction to the configuration parameters 

and apply it before the iteration?

​You can initialize the graph by using any Gelly transformation methods before starting the iteration, e.g. mapVertices, mapEdges, reduceOnEdges, etc.

Btw, a vertex can iterate over its edges inside the ComputeFunction using the getEdges() method. Initializing the vertex values with neighboring edges might not be a good idea if you have vertices with high degrees.​

​Cheers,

-Vasia.​

 

What do you think?

(BTW, I started a PR on FLINK-1526(MST Lib&Example). Considering the complexity, the example is not 

provided.)

Really appreciate for all your help.

Best,

Xingcan

On Thu, Feb 9, 2017 at 5:36 PM, Vasiliki Kalavri <[hidden email]> wrote:

Hi Xingcan,

On 7 February 2017 at 10:10, Xingcan Cui <[hidden email]> wrote:

Hi all,

I got some question about the vertex-centric iteration in Gelly.

a)  It seems the postSuperstep method is called before the superstep barrier (I got different aggregate values of the same superstep in this method). Is this a bug? Or the design is just like that?

​The postSuperstep() method is called inside the close() method of a RichCoGroupFunction that wraps the ComputeFunction. The close() method It is called after the last call to the coGroup() after each iteration superstep.

The aggregate values are not guaranteed to be consistent during the same superstep when they are computed. To retrieve an aggregate value for superstep i, you should use the getPreviousIterationAggregate() method in superstep i+1.

 

b) There is not setHalt method for vertices. When no message received, a vertex just quit the next iteration. Should I manually send messages (like heartbeat) to keep the vertices active?

​That's because vertex halting is implicitly controlled by the underlying delta iterations of Flink. ​A vertex will remain active as long as it receives a message or it updates its value, otherwise it will become inactive. The documentation on Gelly iterations [1] and DataSet iterations [2] might be helpful.

 

c) I think we may need an initialization method in the ComputeFunction.

​There exists a preSuperstep() method for initialization. This one will be executed once per superstep before the compute function is invoked for every vertex. Would this work for you?

 

Any opinions? Thanks.

Best,

Xingcan

​I hope this helps,

-Vasia.​

​[1]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/libs/gelly/iterative_graph_processing.html#vertex-centric-iterations

[2]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/batch/iterations.html​

[Xingcan Cui]

Hi Vasia,

sorry that I should have read the archive before (it's already been posted in FLINK-1526, though with an ugly format). Now everything's clear and I think this thread should be closed here.

Thanks. @Vasia @Greg

Best,

Xingcan

On Tue, Feb 14, 2017 at 3:55 PM, Vasiliki Kalavri <[hidden email]> wrote:

Hi Xingcan,

that's my bad, I was thinking of scatter-gather iterations in my previous reply. You're right, in VertexCentricIteration a vertex is only active in the next superstep if it has received at least one message in the current superstep. Updating its value does not impact the activation. This is intentional in the vertex-centric model.

I agree that the current design of the iterative models is restrictive and doesn't allow for the expression of complex iterative algorithms that require updating edges or defining phases. We have discussed this before, e.g. in [1]. The outcome of that discussion was that we should use for-loop iterations for such cases, as the closed-loop iteration operators of Flink might not provide the necessary flexibility. As you will see in the thread though, that proposal didn't work out, as efficiently supporting for-loops in Flink is not possible right now.

-Vasia.

[1]: http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/DISCUSS-Gelly-iteration-abstractions-td3949.html

On 14 February 2017 at 08:10, Xingcan Cui <[hidden email]> wrote:

Hi Greg,

I also found that in VertexCentricIteration.java, the message set is taken as the workset while the vertex set is taken as the delta for solution set. By doing like that, the setNewVertex method will not actually active a vertex. In other words, if no message is generated (the workset is empty) the "pact.runtime.workset-empty-aggregator" will judge convergence of the delta iteration and then the iteration just terminates. Is this a bug?

Best,

Xingcan

On Mon, Feb 13, 2017 at 5:24 PM, Xingcan Cui <[hidden email]> wrote:

Hi Greg,

Thanks for your attention.

It takes me a little time to read the old PR on FLINK-1885. Though the VertexCentricIteration, as well as its related classes, has been refactored, I understand what Markus want to achieve.

I am not sure if using a bulk iteration instead of a delta one could eliminate the "out of memory" problem.  Except for that, I think the "auto update" has nothing to do with the bulk mode. Considering the compatible guarantee, here is my suggestions to improve gelly's iteration API:

1) Add an "autoHalt" flag to the ComputeFunction.

2) If the flag is set true (default), apply the current mechanism .

3) If the flag is set false, call out.collect() to update the vertex value whether the setNewVertexValue() method is called or not, unless the user explicitly call a (new added) voteToHalt() method in the ComputeFunction.

By adding these, users can decide when to halt a vertex themselves. What do you think? 

As for the "update edge values during vertex iterations" problem, I think it needs a redesign for the gelly framework (Maybe merge the vertices and edges into a single data set? Or just change the iterations'  implementation? I can't think it clearly now.), so that's it for now. Besides, I don't think there will be someone who really would love to write a graph algorithm with Flink native operators and that's why gelly is designed, isn't it?

Best,

Xingcan

On Fri, Feb 10, 2017 at 10:31 PM, Greg Hogan <[hidden email]> wrote:

Hi Xingcan,

FLINK-1885 looked into adding a bulk mode to Gelly's iterative models.

As an alternative you could implement your algorithm with Flink operators and a bulk iteration. Most of the Gelly library is written with native operators.

Greg

On Fri, Feb 10, 2017 at 5:02 AM, Xingcan Cui <[hidden email]> wrote:

Hi Vasia,

b) As I said, when some vertices finished their work in current phase, they have nothing to do (no value updates, no message received, just like slept) but to wait for other vertices that have not finished (the current phase) yet. After that in the next phase, all the vertices should go back to work again and if there are some vertices become inactive in last phase, it could be hard to reactive them again by message since we even don't know which vertices to send to. The only solution is to keep all vertices active, whether by updating vertices values in each super step or sending heartbeat messages to vertices themselves (which will bring a lot of extra work to the MessageCombiner). 

c) I know it's not elegant or even an awful idea to store the edge info into vertex values. However, we can not change edge values or maintain states (even a pick or unpick mark) in edges during a vertex-centric iteration. Then what can we do if an algorithm really need that?

Thanks for your patience.

Best,

Xingcan

On Fri, Feb 10, 2017 at 4:50 PM, Vasiliki Kalavri <[hidden email]> wrote:

Hi Xingcan,

On 9 February 2017 at 18:16, Xingcan Cui <[hidden email]> wrote:

Hi Vasia,

thanks for your reply. It helped a lot and I got some new ideas.

a) As you said, I did use the getPreviousIterationAggregate() method in preSuperstep() of the next superstep.

However, if the (only?) global (aggregate) results can not be guaranteed to be consistency,  what should we

do with the postSuperstep() method?

​The postSuperstep() method is analogous to the close() method in a RichFunction, which is typically used for cleanup.​

 

b) Though we can active vertices by update method or messages, IMO, it may be more proper for users 

themselves to decide when to halt a vertex's iteration. Considering a complex algorithm that contains different 

phases inside a vertex-centric iteration. Before moving to the next phase (that should be synchronized), 

there may be some vertices that already finished their work in current phase and they just wait for others.

Users may choose the finished vertices to idle until the next phase, but rather than to halt them. 

Can we consider adding the voteToHalt() method and some internal variables to the Vertex/Edge class 

(or just create an "advanced" version of them) to make the halting more controllable? 

​I suppose adding a voteToHalt() method is possible, but I'm not sure I see how that would make halting more controllable. If a vertex hasn't changed value or hasn't received a message, it has no work to do in the next iteration, so why keep it active? If in a later superstep, a previously inactive vertex receives a message, it will become active again. ​Is this what you're looking for or am I missing something?

 

c) Sorry that I didn't make it clear before. Here the initialization means a "global" one that executes once 

before the iteration. For example, users may want to initialize the vertices' values by their adjacent edges

before the iteration starts. Maybe we can add an extra coGroupFunction to the configuration parameters 

and apply it before the iteration?

​You can initialize the graph by using any Gelly transformation methods before starting the iteration, e.g. mapVertices, mapEdges, reduceOnEdges, etc.

Btw, a vertex can iterate over its edges inside the ComputeFunction using the getEdges() method. Initializing the vertex values with neighboring edges might not be a good idea if you have vertices with high degrees.​

​Cheers,

-Vasia.​

 

What do you think?

(BTW, I started a PR on FLINK-1526(MST Lib&Example). Considering the complexity, the example is not 

provided.)

Really appreciate for all your help.

Best,

Xingcan

On Thu, Feb 9, 2017 at 5:36 PM, Vasiliki Kalavri <[hidden email]> wrote:

Hi Xingcan,

On 7 February 2017 at 10:10, Xingcan Cui <[hidden email]> wrote:

Hi all,

I got some question about the vertex-centric iteration in Gelly.

a)  It seems the postSuperstep method is called before the superstep barrier (I got different aggregate values of the same superstep in this method). Is this a bug? Or the design is just like that?

​The postSuperstep() method is called inside the close() method of a RichCoGroupFunction that wraps the ComputeFunction. The close() method It is called after the last call to the coGroup() after each iteration superstep.

The aggregate values are not guaranteed to be consistent during the same superstep when they are computed. To retrieve an aggregate value for superstep i, you should use the getPreviousIterationAggregate() method in superstep i+1.

 

b) There is not setHalt method for vertices. When no message received, a vertex just quit the next iteration. Should I manually send messages (like heartbeat) to keep the vertices active?

​That's because vertex halting is implicitly controlled by the underlying delta iterations of Flink. ​A vertex will remain active as long as it receives a message or it updates its value, otherwise it will become inactive. The documentation on Gelly iterations [1] and DataSet iterations [2] might be helpful.

 

c) I think we may need an initialization method in the ComputeFunction.

​There exists a preSuperstep() method for initialization. This one will be executed once per superstep before the compute function is invoked for every vertex. Would this work for you?

 

Any opinions? Thanks.

Best,

Xingcan

​I hope this helps,

-Vasia.​

​[1]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/libs/gelly/iterative_graph_processing.html#vertex-centric-iterations

[2]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/batch/iterations.html​

[Vasiliki Kalavri]

Dear Xingcan,

no need to apologize, we are here to help :) You are always welcome to ask questions / make suggestions.

Cheers,

-Vasia.

On 14 February 2017 at 09:35, Xingcan Cui <[hidden email]> wrote:

Hi Vasia,

sorry that I should have read the archive before (it's already been posted in FLINK-1526, though with an ugly format). Now everything's clear and I think this thread should be closed here.

Thanks. @Vasia @Greg

Best,

Xingcan

On Tue, Feb 14, 2017 at 3:55 PM, Vasiliki Kalavri <[hidden email]> wrote:

Hi Xingcan,

that's my bad, I was thinking of scatter-gather iterations in my previous reply. You're right, in VertexCentricIteration a vertex is only active in the next superstep if it has received at least one message in the current superstep. Updating its value does not impact the activation. This is intentional in the vertex-centric model.

I agree that the current design of the iterative models is restrictive and doesn't allow for the expression of complex iterative algorithms that require updating edges or defining phases. We have discussed this before, e.g. in [1]. The outcome of that discussion was that we should use for-loop iterations for such cases, as the closed-loop iteration operators of Flink might not provide the necessary flexibility. As you will see in the thread though, that proposal didn't work out, as efficiently supporting for-loops in Flink is not possible right now.

-Vasia.

[1]: http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/DISCUSS-Gelly-iteration-abstractions-td3949.html

On 14 February 2017 at 08:10, Xingcan Cui <[hidden email]> wrote:

Hi Greg,

I also found that in VertexCentricIteration.java, the message set is taken as the workset while the vertex set is taken as the delta for solution set. By doing like that, the setNewVertex method will not actually active a vertex. In other words, if no message is generated (the workset is empty) the "pact.runtime.workset-empty-aggregator" will judge convergence of the delta iteration and then the iteration just terminates. Is this a bug?

Best,

Xingcan

On Mon, Feb 13, 2017 at 5:24 PM, Xingcan Cui <[hidden email]> wrote:

Hi Greg,

Thanks for your attention.

It takes me a little time to read the old PR on FLINK-1885. Though the VertexCentricIteration, as well as its related classes, has been refactored, I understand what Markus want to achieve.

I am not sure if using a bulk iteration instead of a delta one could eliminate the "out of memory" problem.  Except for that, I think the "auto update" has nothing to do with the bulk mode. Considering the compatible guarantee, here is my suggestions to improve gelly's iteration API:

1) Add an "autoHalt" flag to the ComputeFunction.

2) If the flag is set true (default), apply the current mechanism .

3) If the flag is set false, call out.collect() to update the vertex value whether the setNewVertexValue() method is called or not, unless the user explicitly call a (new added) voteToHalt() method in the ComputeFunction.

By adding these, users can decide when to halt a vertex themselves. What do you think? 

As for the "update edge values during vertex iterations" problem, I think it needs a redesign for the gelly framework (Maybe merge the vertices and edges into a single data set? Or just change the iterations'  implementation? I can't think it clearly now.), so that's it for now. Besides, I don't think there will be someone who really would love to write a graph algorithm with Flink native operators and that's why gelly is designed, isn't it?

Best,

Xingcan

On Fri, Feb 10, 2017 at 10:31 PM, Greg Hogan <[hidden email]> wrote:

Hi Xingcan,

FLINK-1885 looked into adding a bulk mode to Gelly's iterative models.

As an alternative you could implement your algorithm with Flink operators and a bulk iteration. Most of the Gelly library is written with native operators.

Greg

On Fri, Feb 10, 2017 at 5:02 AM, Xingcan Cui <[hidden email]> wrote:

Hi Vasia,

b) As I said, when some vertices finished their work in current phase, they have nothing to do (no value updates, no message received, just like slept) but to wait for other vertices that have not finished (the current phase) yet. After that in the next phase, all the vertices should go back to work again and if there are some vertices become inactive in last phase, it could be hard to reactive them again by message since we even don't know which vertices to send to. The only solution is to keep all vertices active, whether by updating vertices values in each super step or sending heartbeat messages to vertices themselves (which will bring a lot of extra work to the MessageCombiner). 

c) I know it's not elegant or even an awful idea to store the edge info into vertex values. However, we can not change edge values or maintain states (even a pick or unpick mark) in edges during a vertex-centric iteration. Then what can we do if an algorithm really need that?

Thanks for your patience.

Best,

Xingcan

On Fri, Feb 10, 2017 at 4:50 PM, Vasiliki Kalavri <[hidden email]> wrote:

Hi Xingcan,

On 9 February 2017 at 18:16, Xingcan Cui <[hidden email]> wrote:

Hi Vasia,

thanks for your reply. It helped a lot and I got some new ideas.

a) As you said, I did use the getPreviousIterationAggregate() method in preSuperstep() of the next superstep.

However, if the (only?) global (aggregate) results can not be guaranteed to be consistency,  what should we

do with the postSuperstep() method?

​The postSuperstep() method is analogous to the close() method in a RichFunction, which is typically used for cleanup.​

 

b) Though we can active vertices by update method or messages, IMO, it may be more proper for users 

themselves to decide when to halt a vertex's iteration. Considering a complex algorithm that contains different 

phases inside a vertex-centric iteration. Before moving to the next phase (that should be synchronized), 

there may be some vertices that already finished their work in current phase and they just wait for others.

Users may choose the finished vertices to idle until the next phase, but rather than to halt them. 

Can we consider adding the voteToHalt() method and some internal variables to the Vertex/Edge class 

(or just create an "advanced" version of them) to make the halting more controllable? 

​I suppose adding a voteToHalt() method is possible, but I'm not sure I see how that would make halting more controllable. If a vertex hasn't changed value or hasn't received a message, it has no work to do in the next iteration, so why keep it active? If in a later superstep, a previously inactive vertex receives a message, it will become active again. ​Is this what you're looking for or am I missing something?

 

c) Sorry that I didn't make it clear before. Here the initialization means a "global" one that executes once 

before the iteration. For example, users may want to initialize the vertices' values by their adjacent edges

before the iteration starts. Maybe we can add an extra coGroupFunction to the configuration parameters 

and apply it before the iteration?

​You can initialize the graph by using any Gelly transformation methods before starting the iteration, e.g. mapVertices, mapEdges, reduceOnEdges, etc.

Btw, a vertex can iterate over its edges inside the ComputeFunction using the getEdges() method. Initializing the vertex values with neighboring edges might not be a good idea if you have vertices with high degrees.​

​Cheers,

-Vasia.​

 

What do you think?

(BTW, I started a PR on FLINK-1526(MST Lib&Example). Considering the complexity, the example is not 

provided.)

Really appreciate for all your help.

Best,

Xingcan

On Thu, Feb 9, 2017 at 5:36 PM, Vasiliki Kalavri <[hidden email]> wrote:

Hi Xingcan,

On 7 February 2017 at 10:10, Xingcan Cui <[hidden email]> wrote:

Hi all,

I got some question about the vertex-centric iteration in Gelly.

a)  It seems the postSuperstep method is called before the superstep barrier (I got different aggregate values of the same superstep in this method). Is this a bug? Or the design is just like that?

​The postSuperstep() method is called inside the close() method of a RichCoGroupFunction that wraps the ComputeFunction. The close() method It is called after the last call to the coGroup() after each iteration superstep.

The aggregate values are not guaranteed to be consistent during the same superstep when they are computed. To retrieve an aggregate value for superstep i, you should use the getPreviousIterationAggregate() method in superstep i+1.

 

b) There is not setHalt method for vertices. When no message received, a vertex just quit the next iteration. Should I manually send messages (like heartbeat) to keep the vertices active?

​That's because vertex halting is implicitly controlled by the underlying delta iterations of Flink. ​A vertex will remain active as long as it receives a message or it updates its value, otherwise it will become inactive. The documentation on Gelly iterations [1] and DataSet iterations [2] might be helpful.

 

c) I think we may need an initialization method in the ComputeFunction.

​There exists a preSuperstep() method for initialization. This one will be executed once per superstep before the compute function is invoked for every vertex. Would this work for you?

 

Any opinions? Thanks.

Best,

Xingcan

​I hope this helps,

-Vasia.​

​[1]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/libs/gelly/iterative_graph_processing.html#vertex-centric-iterations

[2]: https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/batch/iterations.html​