[
https://issues.apache.org/jira/browse/IGNITE-2265?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15071248#comment-15071248
]
Vladimir Ozerov edited comment on IGNITE-2265 at 12/24/15 9:11 PM:
-------------------------------------------------------------------
Andrey,
You are asking tough questions :-) It is hard to reason whether particular
microoptimziation will give any observable benefit or not.
Basically, I am trying to find places which could speedup the product. Clearly,
there are lot of them. So for now I am looking for suspicious code pieces which
are both a) easy to improve; b) have good chances to increase overall
performance.
To identify them I simply run base scenarios like GET/PUT and see profiler
results in terms of hot execution paths and memory allocations.
Atomic* variables is one of the hotspots revealed by profiler. Some atomics are
located inside our cache futures which are created on virually any cache
operation. And in a simple benchmark with distributed cache GET, these Atomics*
are responsible for 2-3% of all heap allocations We can easily get rid of
these Atomics and decrease memory pressure.
2-3% is not a big number when considered separately. But if more similar
optimizations are applied, we could reduce memory pressure in common scenarios
by about 20-25% with minimal efforts. I log such potentially promising finding
as subtasks to the parent ticket IGNITE-2232.
was (Author: vozerov):
Andrey,
You are asking tough questions :-) It is hard to reason whether particular
microoptimziation will give any observable benefit or not.
Basically, I am trying to find places which could speedup the product. Clearly,
there are lot of them. So for now I am looking for suspicious code pieces which
are both a) easy to improve; b) have good chances to increase overall
performance.
To identify them I simply run base scenarios like GET/PUT and see profiler
results in terms of hot execution paths and memory allocations.
Atomic* variables is one of the hotspots revealed by profiler. Some atomics are
located inside our cache futures which are created on virually any cache
operation. And in a simple benchmark with distributed cache GET, these Atomics*
are responsible for 2-3% of all heap allocations We can easily get rid of
these Atomics and decrease memory pressure.
2-3% is not a big number when considered separately. But if more similar
optimizations are applied, we could reduce memory pressure in common scenarios
by about 20-25% with minimal efforts. I log such potentially promising finding
as subtask to the parent ticket IGNITE-2232.
> Replace Atomic* variables with field updaters.
> ----------------------------------------------
>
> Key: IGNITE-2265
> URL: https://issues.apache.org/jira/browse/IGNITE-2265
> Project: Ignite
> Issue Type: Sub-task
> Components: general
> Affects Versions: ignite-1.4
> Reporter: Vladimir Ozerov
> Assignee: Vladimir Ozerov
> Fix For: 1.6
>
>
> Field updaters must be used instaed of the following classes:
> 1) AtomicInteger
> 2) AtomicLong
> 3) AtomicReference
> 4) If AtomicBoolean is met in some hot spaces, it must be replaced with (int
> + updater) pair.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
