Re: Asynchronous I/O poor performance

Posted by Arvid Heise-3 on
URL: http://deprecated-apache-flink-user-mailing-list-archive.369.s1.nabble.com/Fwd-Asynchronous-I-O-poor-performance-tp36382p36430.html

Hi Mark,

could you please check if you can tune akka? Usually in async I/O, the used library uses a thread pool that becomes the actual bottleneck.

If you configure async I/O to use a capacity of 100 and parallelism of 8 on one node, you also need to have ~800 threads in akka (500 might be enough because of overhead) or else async I/O gets blocked while waiting for akka threads to become available.

Best,

Arvid

On Mon, Jul 6, 2020 at 1:45 PM Benchao Li <[hidden email]> wrote:

Hi Mark,

According to your data, I think the config of AsyncOperator is OK.

There is one more config that might affect the throughput of AsyncOperator, it's watermark.

Because unordered async operator still keeps the order between watermarks, did you use

event time in your job, and if yes, what's the watermark interval in your job?

Mark Zitnik <[hidden email]> 于2020年7月5日周日 下午7:44写道：

Hi Benchao

The capacity is 100

Parallelism is 8 

Rpc req is 20ms 

Thanks

On Sun, 5 Jul 2020, 6:16 Benchao Li, <[hidden email]> wrote:

Hi Mark,

Could you give more details about your Flink job?

- the capacity of AsyncDataStream

- the parallelism of AsyncDataStream operator

- the time of per blocked rpc request

Mark Zitnik <[hidden email]> 于2020年7月5日周日 上午3:48写道：

Hi 

 

In my flink application I need to enrich data using AsyncDataStream.unorderedWait but I am getting poor perforce at the beginning I was just working with http call, but I have switched to grpc, I running on 8 core node and getting total of 3200 events per second my service that I am using is not fully utilized and can produce up to 10000 req/seq

Flink job flow 

Reading from Kafka ~> some enrichment with unoderedwait ~> map ~> write to Kafka 

Using Akkad grpc code written in scala 

Thanks

--

Best,
Benchao Li

--

Best,
Benchao Li

--