[
https://issues.apache.org/jira/browse/STORM-297?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14010859#comment-14010859
]
ASF GitHub Bot commented on STORM-297:
--------------------------------------
Github user Gvain commented on the pull request:
https://github.com/apache/incubator-storm/pull/103#issuecomment-44372134
@clockfly
I used latest storm to reproduce some of your test. I found that adding a
few more workers will increase throughput, as well as CPU usage. I also get a
max doubled throughput.
Im my test, i used 4 nodes(24 cores each, not power enough than yours), 48
spouts, 48 bolts, 48 ackers, and 4 to 64 workers, SOL benchmark, message size
is 100 Bytes. Here is the results:
workers | Throughput | CPU usage | NET usage (only IN Bytes)
---------------------------------------------------------------------
4 | 320,000 tps | 56% | 14MB/s
8 | 656,000 tps | 89% | 28MB/s
16 | 560,000 tps | 92% | 26MB/s
32 | 353,000 tps | 90% | 20MB/s
64 | 208,000 tps | 90% | 16MB/s
Using 8 workers, the throughput doubled.
As we already discussed:
(a) Increasing worker count, not too much, may not cause netty context
switching problem. As using a SHARED threadpool(its default size is 1) among
all netty client within a worker, the netty threads number do not increase as
total worker numbers increase. Check
https://github.com/apache/incubator-storm/pull/57
(b) Increasing worker count will increase netty threads. Using more netty
threads working in sync and non-batch mode may have some what the same effect
with using less netty threads working in async and batch mode which is your way.
From the test result, it seems like the above point of view (b) is true.
So, maybe, what we need to to is just pulling the
https://github.com/apache/incubator-storm/pull/57 and adding a few workers.
But i am still curious about if i adding more nodes, should this still work
?
> Storm Performance cannot be scaled up by adding more CPU cores
> --------------------------------------------------------------
>
> Key: STORM-297
> URL: https://issues.apache.org/jira/browse/STORM-297
> Project: Apache Storm (Incubating)
> Issue Type: Bug
> Reporter: Sean Zhong
> Labels: Performance, netty
> Fix For: 0.9.2-incubating
>
> Attachments: Storm_performance_fix.pdf,
> storm_Netty_receiver_diagram.png, storm_performance_fix.patch
>
>
> We cannot scale up the performance by adding more CPU cores and increasing
> parallelism.
> For a 2 layer topology Spout ---shuffle grouping--> bolt, when message size
> is small (around 100 bytes), we can find in the below picture that neither
> the CPU nor the network is saturated. When message size is 100 bytes, only
> 40% of CPU is used, only 18% of network is used, although we have a high
> parallelism (overall we have 144 executors)
--
This message was sent by Atlassian JIRA
(v6.2#6252)
