[jkaczman]

Previously, we were undergoing a lot of unnecessary overhead by awaiting on the admin termination (FORCE_RELOAD) and shutdown cancellation futures WITHIN (on every iteration of) our Client::run loop.

This reduces the overhead by placing these awaits on our cancellation token futures within a select! above the loop, so that we only await these futures a singular time per instance of Client::run (vs. every loop iteration).

Local flamegraph shows this reduces our total overhead by about 1-2%.

Arm64 and x86 benchmarks show a 4% increase in TPS when dealing with 8 workers.

[codecov]

Codecov Report

❌ Patch coverage is 93.10345% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
pgdog/src/frontend/client/mod.rs	91.66%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!

[github-actions]

Metal bench (arm64) passed. All 148 measures within bounds.

suite	throughput	cpu/txn	user cycles/txn	user instr/txn	kernel cycles/txn	ctx/txn
pooler		26.26 (-0.2% ± 0.8%)	26,147 (-1.0% ± 0.8%)	34,808 (-1.1% ± 0.02%)	38,096 (+0.7% ± 1.3%)
lb		27.06 (-1.0% ± 1.3%)	27,352 (-2.0% ± 1.1%)	35,681 (-1.0% ± 0.03%)	38,652 (-0.4% ± 1.7%)
sharding-w2		48.48 (-2.7%)	62,140 (-3.1%)	70,975 (-1.8%)	58,013 (-2.3%)	1.397 (+0.3%)
sharding-w8	165,399 (+3.9%)	43.77 (-4.8%)	62,456 (-6.2%)	54,846 (-2.2%)	50,356 (-3.1%)	0.1624 (+8.1%)

Counters per txn

suite	frontend stall	l1i refill	backend stall
pooler	12,109 (-0.2% ± 1.8%)	3,033 (-0.7% ± 0.7%)	2,722 (-4.7% ± 1.3%)
lb	12,742 (-2.6% ± 2.4%)	3,119 (-1.3% ± 0.8%)	2,996 (-3.2% ± 1.7%)
sharding-w2	34,299 (-4.0%)	4,996 (-3.6%)	6,815 (+1.0%)
sharding-w8	34,289 (-9.3%)	4,732 (-4.9%)	9,412 (+0.3%)

Microbench: all 135 measures within bounds.

lb flamegraph vs 97fc159

Open interactive for hover, zoom and search.

Open interactive.

microbench flamegraph vs 97fc159

Open interactive for hover, zoom and search.

Open interactive.

pooler flamegraph vs 97fc159

Open interactive for hover, zoom and search.

Open interactive.

sharding-w2 flamegraph vs 97fc159

Open interactive for hover, zoom and search.

Open interactive.

sharding-w8 flamegraph vs 97fc159

Open interactive for hover, zoom and search.

Open interactive.

Logs

lb, microbench, pooler, sharding-w2, sharding-w8, verdict.json

Compared d75b48c with 97fc159 on c7g.metal (on-demand).

[github-actions]

Metal bench (x86) failed. 1 of 148 measures outside bounds.

suite	throughput	cpu/txn	user cycles/txn	user instr/txn	kernel cycles/txn	ctx/txn
pooler		25.92 (-0.8% ± 0.2%)	19,922 (-2.8% ± 0.1%)	36,447 (-0.9% ± 0.02%)	33,507 (+0.2% ± 0.4%)
lb		26.29 (-0.9% ± 0.1%)	20,510 (-2.7% ± 0.1%)	37,392 (-0.9% ± 0.02%)	33,575 (-0.0% ± 0.3%)
sharding-w2		40.25 (-2.7%)	43,581 (-3.5%)	74,518 (-1.7%)	39,320 (-3.0%)	1.519 (+0.4%)
sharding-w8	176,865 (+3.9%)	38.03 (-2.5%)	48,100 (-2.1%)	57,707 (-2.3%)	41,501 (-3.1%)	0.1865 (-0.1%)

Counters per txn

suite	frontend stall	l1i refill	backend stall
pooler	3,222 (-7.2% ± 0.3%)	1,133 (-3.4% ± 0.5%)	41,667 (-2.0% ± 0.2%)
lb	3,323 (-6.6% ± 0.4%)	1,149 (-3.8% ± 0.5%)	43,167 (-2.1% ± 0.4%)
sharding-w2	5,668 (-11.7%)	1,610 (-5.4%)	119,262 (-1.4%)
sharding-w8	3,213 (-8.7%)	972 (-4.0%)	183,988 (-0.9%)

benchmark	measure	result	baseline	change	bound
extended_cached::delete_by_key	latency	2,875	2,727	+5.42% ± 0.93%	+5%

Microbench: 134 of 135 measures within bounds.

• extended_cached: 15 of 15 benches slower, median +3.0% ± 0.8%
• extended_cold: 13 of 15 benches slower, median +1.4% ± 0.7%
• simple_cold: 13 of 15 benches slower, median +1.7% ± 0.5%

lb flamegraph vs 97fc159

Open interactive for hover, zoom and search.

Open interactive.

microbench flamegraph vs 97fc159

Open interactive for hover, zoom and search.

Open interactive.

pooler flamegraph vs 97fc159

Open interactive for hover, zoom and search.

Open interactive.

sharding-w2 flamegraph vs 97fc159

Open interactive for hover, zoom and search.

Open interactive.

sharding-w8 flamegraph vs 97fc159

Open interactive for hover, zoom and search.

Open interactive.

Logs

lb, microbench, pooler, sharding-w2, sharding-w8, verdict.json

Compared d75b48c with 97fc159 on c7i.metal-48xl (spot).