[
https://issues.apache.org/jira/browse/IGNITE-29099?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Nikita Amelchev updated IGNITE-29099:
-------------------------------------
Description:
Every cache update boxes primitive map keys on its hot path:
* \{{GridCacheSharedContext#cacheContext(int)}} and
\{{#cacheObjectContext(int)}} look up a \{{ConcurrentHashMap<Integer,
GridCacheContext>}}. A cache ID is a hash of the cache name, so it misses the
\{{Integer}} cache and a new object is allocated on each call. The methods are
called several times per operation on every node involved.
* \{{TcpCommunicationMetricsListener#updateMessageTypeMap}} checks a
\{{Map<Short, String>}} for every sent and received message.
These allocations are pure garbage produced on NIO and striped pool threads of
all nodes. Both lookups should use primitive keys.
was:
JFR profiling of a FULL_SYNC transactional put (3 server nodes, 2 backups)
shows three allocation sites that box a primitive map key on every call:
* {{GridCacheSharedContext#cacheContext(int)}} and {{#cacheObjectContext(int)}}
- lookup in {{ConcurrentHashMap<Integer, GridCacheContext>}}. A cache ID is a
hash of the cache name, so it misses the {{Integer}} cache. Called several
times per transaction on each node: {{IgniteTxEntryMarshaller}},
{{IgniteTxEntry#initializeContext}}, {{IgniteTxStateImpl#validateTopology}},
{{IgniteTxStateImpl#onTxEnd}}.
* {{TcpCommunicationMetricsListener#updateMessageTypeMap}} - {{Map<Short,
String>#containsKey(short)}} for every sent and received message.
Together these sites produce 2-3% of all allocations of a transaction.
h3. Proposed change
* {{GridCacheSharedContext}}: add a volatile copy-on-write
{{IntMap<GridCacheContext>}} snapshot of {{ctxMap}}, rebuilt in
{{addCacheContext}} and {{removeCacheContext}}; both lookups read the snapshot.
* {{TcpCommunicationMetricsListener}}: replace {{Map<Short, String>}} with a
copy-on-write {{IntMap<String>}}.
h3. Result
JMH, {{gc.alloc.rate.norm}}, implicit put: the three sites disappear from the
JFR profile, allocations per transaction drop by about 1.7% (28.6 KB -> 28.1
KB), throughput is unchanged.
{{JmhCacheTxFullSyncBenchmark}} is added to reproduce the measurements.
> Avoid boxing of cache ID and message direct type on the cache operation hot
> path
> --------------------------------------------------------------------------------
>
> Key: IGNITE-29099
> URL: https://issues.apache.org/jira/browse/IGNITE-29099
> Project: Ignite
> Issue Type: Improvement
> Reporter: Nikita Amelchev
> Assignee: Nikita Amelchev
> Priority: Minor
> Attachments: JmhCacheTxFullSyncBenchmark.java
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> Every cache update boxes primitive map keys on its hot path:
> * \{{GridCacheSharedContext#cacheContext(int)}} and
> \{{#cacheObjectContext(int)}} look up a \{{ConcurrentHashMap<Integer,
> GridCacheContext>}}. A cache ID is a hash of the cache name, so it misses the
> \{{Integer}} cache and a new object is allocated on each call. The methods
> are called several times per operation on every node involved.
> * \{{TcpCommunicationMetricsListener#updateMessageTypeMap}} checks a
> \{{Map<Short, String>}} for every sent and received message.
> These allocations are pure garbage produced on NIO and striped pool threads
> of all nodes. Both lookups should use primitive keys.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
