Refactor Client Write Preparation and Handling

[fcostaoliveira]

This update refactors prepareClientToWrite by introducing _prepareClientToWrite for inline checks within network.c file, and separates replica and non-replica handling for pending replies and writes (_clientHasPendingRepliesSlave/NonSlave and _writeToClientSlave/NonSlave).

prepareClientToWrites improvement

Function / Call Stack	CPU Time: Difference	CPU Time: unstable (04f63d4)	CPU Time: This PR (0584dfe)
prepareClientToWrite	2.40%	3.55%	1.15%

splitting replica and non-replica handling improvement

TBD

To test it we can simply:

tclsh tests/test_helper.tcl --single unit/type/list

Confirmation from daily:

• https://github.com/filipecosta90/redis/actions/runs/12653274807

benchmark results

Achievable ops/sec impact:

test name	unstable (04f63d4)	This PR first commit (22b75ff)	% change	This PR last commit (145de69)	% change
memtier_benchmark-1key-list-1K-elements-lrange-all-elements	14746	15253	3.4%	15607	5.8%
memtier_benchmark-1key-list-100-elements-lrange-all-elements	82602	80676	-2.3%	85011	2.9%
memtier_benchmark-1key-list-100-elements-int-7bit-uint-lrange-all-elements-pipeline-10	136793	139121	1.7%	148916	8.9%
memtier_benchmark-1key-list-100-elements-lrange-all-elements-pipeline-10	137138	135291	-1.3%	141980	3.5%
memtier_benchmark-1key-list-1K-elements-lrange-all-elements-pipeline-10	13975	14256	2.0%	14798	5.9%
memtier_benchmark-1key-list-2K-elements-quicklist-lrange-all-elements-longs	5996	6239	4.1%	6330	5.6%
memtier_benchmark-1key-list-10-elements-lrange-all-elements-pipeline-10	518595	522612	0.8%	536865	3.5%
memtier_benchmark-1key-list-100-elements-int-lrange-all-elements-pipeline-10	109166	110532	1.3%	113514	4.0%
memtier_benchmark-1key-list-10-elements-lrange-all-elements	145262	143343	-1.3%	144856	-0.3%
memtier_benchmark-1key-hash-1K-fields-hgetall	7407.93	7821.4	5.6%	8025.5	8.3%
memtier_benchmark-1key-hash-1K-fields-hgetall-pipeline-10	7156.6	7409.02	3.5%	7650.24	6.9%

To reproduce:

taskset -c 0 /root/redis/src/redis-server  --unixsocket /tmp/1.socket --save '' --enable-debug-command local
pip3 install redis-benchmarks-specification
redis-benchmarks-spec-client-runner --tests-regexp ".*lrange.*|memtier_benchmark-1key-hash-1K-fields-hgetall.*" --flushall_on_every_test_start --flushall_on_every_test_end  --cpuset_start_pos 2 --override-memtier-test-time 60

[sundb]

fully CI: https://github.com/redis/redis/actions/runs/12671077818

[fcostaoliveira]

CE Performance Automation : step 2 of 2 (benchmark) FINISHED.

This comment was automatically generated given a benchmark was triggered.

Started benchmark suite at 2025-01-10 19:31:08.441028 and took 8152.022405 seconds to finish.
Status: [################################################################################] 100.0% completed.

In total will run 135 benchmarks.
- 0 pending.
- 135 completed:
- 4 successful.
- 131 failed.
You can check a the status in detail via the grafana link

[fcostaoliveira]

Automated performance analysis summary

This comment was automatically generated given there is performance data available.

Using platform named: intel64-ubuntu22.04-redis-clx1 to do the comparison.

In summary:

• Detected a total of 4 improvements above the improvement water line.

You can check a comparison in detail via the grafana link

Comparison between unstable and filipecosta90:optimize.prepareClientToWrite.

Time Period from 5 months ago. (environment used: oss-standalone)

Improvements Table

Test Case	Baseline redis/redis unstable (median obs. +- std.dev)	Comparison redis/redis filipecosta90:optimize.prepareClientToWrite (median obs. +- std.dev)	% change (higher-better)	Note
memtier_benchmark-10Kkeys-load-hash-50-fields-with-10000B-values	1084	1823	68.2%	IMPROVEMENT
memtier_benchmark-1Mkeys-generic-scan-pipeline-10	249738	322217	29.0%	IMPROVEMENT
memtier_benchmark-1Mkeys-load-stream-5-fields-with-100B-values-pipeline-10	95311	139362	46.2%	IMPROVEMENT
memtier_benchmark-2keys-zset-300-elements-skiplist-encoded-zunion	1950	2506	28.5%	IMPROVEMENT

Improvements test regexp names: memtier_benchmark-10Kkeys-load-hash-50-fields-with-10000B-values|memtier_benchmark-1Mkeys-generic-scan-pipeline-10|memtier_benchmark-1Mkeys-load-stream-5-fields-with-100B-values-pipeline-10|memtier_benchmark-2keys-zset-300-elements-skiplist-encoded-zunion

Full Results table:

Test Case	Baseline redis/redis unstable (median obs. +- std.dev)	Comparison redis/redis filipecosta90:optimize.prepareClientToWrite (median obs. +- std.dev)	% change (higher-better)	Note
memtier_benchmark-10Kkeys-load-hash-50-fields-with-10000B-values	1084	1823	68.2%	IMPROVEMENT
memtier_benchmark-1Mkeys-generic-scan-pipeline-10	249738	322217	29.0%	IMPROVEMENT
memtier_benchmark-1Mkeys-load-stream-5-fields-with-100B-values-pipeline-10	95311	139362	46.2%	IMPROVEMENT
memtier_benchmark-2keys-zset-300-elements-skiplist-encoded-zunion	1950	2506	28.5%	IMPROVEMENT

[ShooterIT]

Hi @fcostaoliveira and do you know if __attribute__((always_inline)) can work when we add it at the beginning of a normal function?

[ShooterIT]

I tried __attribute__((always_inline)) int prepareClientToWrite(client *c), gcc in ubuntu reports a warning networking.c:286:36: warning: 'always_inline' function might not be inlinable [-Wattributes] , but clang in mac didn't.
it seems _attribute__((always_inline)) is recommended to be used together with inline. I initially want to just use __attribute__((always_inline)) int prepareClientToWrite(client *c) to avoid changing much code and dirtying blame log, but it doesn't work, so we can leave as it is