[jkaczman]

Follow up work to try to improve accuracy of microbench instructions in reference to #2003 seeing a +1.54% increase for a measure that shouldn't have been affected by that PR.

After some investigation, this seems to be because our LRU caches hash with foldhash, which builds a hash seed from code addresses (as a source of randomness). My guess is that this can change the order that we free cache entries in, and with that differs the work of all subsequent malloc calls.

Reference to their code that does this: https://github.com/orlp/foldhash/blob/77d8e3dae0d43abe65753e888b9e74efbe4ee3f7/src/seed.rs#L150-L156

Now, in the microbench, the LRU caches hash with a fixed seed (via foldhash::fast::FixedState, which they recommend for determinism).

This doesn't affect production builds.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[levkk]

You could make it conditional, i.e., only if bench feature is enabled. Although LLVM will probably remove it anyway during compilation / dead code elimination.

[github-actions]

Metal bench (arm64) passed. All 517 measures within bounds.

suite	throughput	cpu/txn	user cycles/txn	user instr/txn	kernel cycles/txn	kernel instr/txn	ctx/txn
pooler		19.54 (-0.3% ± 1.4%)	19,370 (-0.8% ± 0.6%)	32,042 (+0.0% ± 0.07%)	30,905 (-0.2% ± 2.5%)	50,975 (-0.1% ± 0.3%)
lb		20.12 (-0.2% ± 2.2%)	20,112 (+0.7% ± 1.5%)	32,939 (-0.0% ± 0.05%)	31,382 (-0.9% ± 3.2%)	52,133 (-0.0% ± 0.4%)
sharding-w2		32.98 (-0.7% ± 1.7%)	44,777 (-0.5% ± 1.2%)	97,448 (+0.0% ± 0.00%)	41,220 (-1.0% ± 4.3%)	53,628 (+0.0% ± 0.2%)	1.54 (+0.1% ± 0.2%)
sharding-w8	243,019 (+2.1% ± 3.2%)	29.48 (-2.3% ± 2.9%)	44,674 (-1.6% ± 1.4%)	52,985 (-0.0% ± 0.1%)	36,790 (-3.3% ± 5.3%)	42,384 (+0.0% ± 0.07%)	0.1675 (+2.6% ± 1.3%)

Counters per txn

suite	frontend stall	l1i refill	backend stall
pooler	9,043 (-2.4% ± 1.6%)	2,814 (-1.3% ± 1.0%)
lb	9,547 (+0.7% ± 3.5%)	2,899 (+0.1% ± 1.6%)
sharding-w2	19,151 (-1.0% ± 3.3%)	4,231 (-0.6% ± 2.7%)
sharding-w8	20,802 (-3.0% ± 2.3%)	4,191 (-1.0% ± 2.5%)

Microbench: all 504 measures within bounds.

Hot code per txn

path	base	head	change
pooler	111.2 KB	110.6 KB	-0.6 KB
sharding-w2	166.8 KB	166.8 KB	+0.1 KB

pooler:

crate	base	head	change
pgdog::frontend::client::query_engine	19.8 KB	19.8 KB	+0.0 KB
tokio	18.4 KB	18.2 KB	-0.2 KB
pgdog::backend::pool	17.7 KB	17.4 KB	-0.2 KB
aws-lc (TLS crypto)	10.0 KB	9.8 KB	-0.2 KB
rustls (TLS)	8.4 KB	8.4 KB	+0.1 KB
pgdog::frontend (other)	7.8 KB	7.8 KB	+0.0 KB
pgdog::backend::server	7.2 KB	7.3 KB	+0.1 KB
other	6.4 KB	6.3 KB	-0.1 KB
everything else	15.5 KB	15.5 KB	+0.0 KB

function	head	change
pgdog::frontend::client::Client::run::{{closure}}	5.4 KB	-0.1 KB
pgdog::frontend::client::query_engine::query::<impl pgdog::frontend::client::query_engine:	5.4 KB	+0.0 KB
pgdog::frontend::client::query_engine::QueryEngine::handle::{{closure}}	4.0 KB	+0.0 KB
tokio::runtime::task::raw::poll	3.0 KB	-0.2 KB
pgdog::frontend::client::query_engine::query::<impl pgdog::frontend::client::query_engine:	2.8 KB	+0.0 KB

sharding-w2:

crate	base	head	change
pgdog::frontend::router	29.0 KB	28.8 KB	-0.2 KB
pgdog::frontend::client::query_engine	20.9 KB	20.9 KB	+0.0 KB
tokio	19.4 KB	19.4 KB	+0.0 KB
pgdog::backend::pool	18.5 KB	18.6 KB	+0.1 KB
aws-lc (TLS crypto)	12.5 KB	13.0 KB	+0.5 KB
other	9.6 KB	9.9 KB	+0.4 KB
pgdog::frontend (other)	8.7 KB	8.6 KB	-0.1 KB
rustls (TLS)	8.2 KB	8.2 KB	+0.0 KB
everything else	40.0 KB	39.5 KB	-0.5 KB

function	head	change
pgdog::frontend::client::Client::run::{{closure}}	6.0 KB	-0.1 KB
pgdog::frontend::client::query_engine::query::<impl pgdog::frontend::client::query_engine:	5.1 KB	-0.1 KB
pgdog::frontend::client::query_engine::QueryEngine::handle::{{closure}}	4.1 KB	+0.0 KB
pgdog::frontend::router::parser::query::select::<impl pgdog::frontend::router::parser::que	3.4 KB	+0.1 KB
tokio::runtime::task::raw::poll	3.3 KB	+0.0 KB

lb flamegraph vs e9316d2

Open interactive for hover, zoom and search.

Open interactive.

microbench flamegraph vs e9316d2

Open interactive for hover, zoom and search.

Open interactive.

pooler flamegraph vs e9316d2

Open interactive for hover, zoom and search.

Open interactive.

sharding-w2 flamegraph vs e9316d2

Open interactive for hover, zoom and search.

Open interactive.

sharding-w8 flamegraph vs e9316d2

Open interactive for hover, zoom and search.

Open interactive.

Logs

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

Compared a2b6aa9 with e9316d2 on m8g.metal-24xl (spot).

[github-actions]

Metal bench (x86) passed. All 517 measures within bounds.

suite	throughput	cpu/txn	user cycles/txn	user instr/txn	kernel cycles/txn	kernel instr/txn	ctx/txn
pooler		34.63 (+0.1% ± 0.2%)	22,808 (+0.5% ± 0.4%)	33,987 (-0.0% ± 0.02%)	57,821 (+0.1% ± 0.4%)	38,540 (-0.1% ± 0.1%)
lb		35.58 (-0.0% ± 0.2%)	23,853 (+0.3% ± 0.4%)	34,935 (-0.0% ± 0.02%)	58,755 (-0.1% ± 0.3%)	39,233 (+0.0% ± 0.1%)
sharding-w2		56.63 (+0.2% ± 0.5%)	53,668 (+0.6% ± 0.7%)	101,372 (-0.0% ± 0.06%)	73,140 (-0.0% ± 0.9%)	43,130 (-0.0% ± 0.4%)	1.392 (-0.2% ± 0.3%)
sharding-w8	148,661 (-0.0% ± 1.5%)	48.35 (+0.4% ± 0.2%)	58,552 (+0.8% ± 1.1%)	56,812 (+0.0% ± 0.07%)	64,317 (+0.1% ± 0.5%)	34,215 (-0.0% ± 0.2%)	0.1673 (+0.0% ± 2.0%)

Counters per txn

suite	frontend stall	l1i refill	backend stall
pooler	72,760 (+1.4% ± 0.6%)	3,606 (+1.5% ± 0.3%)
lb	75,668 (+1.3% ± 1.1%)	3,733 (+1.5% ± 0.2%)
sharding-w2	161,448 (+1.6% ± 1.1%)	6,008 (+1.0% ± 0.7%)
sharding-w8	127,138 (+1.9% ± 0.4%)	4,868 (+0.8% ± 1.3%)

Microbench: all 504 measures within bounds.

Hot code per txn

path	base	head	change
pooler	103.9 KB	105.6 KB	+1.7 KB
sharding-w2	158.9 KB	160.4 KB	+1.4 KB

pooler:

crate	base	head	change
pgdog::frontend::client::query_engine	19.6 KB	20.1 KB	+0.5 KB
tokio	17.4 KB	17.6 KB	+0.2 KB
pgdog::backend::pool	17.4 KB	17.4 KB	+0.0 KB
pgdog::frontend (other)	8.6 KB	9.0 KB	+0.4 KB
rustls (TLS)	6.9 KB	7.8 KB	+0.9 KB
aws-lc (TLS crypto)	7.3 KB	7.0 KB	-0.3 KB
pgdog::backend::server	6.4 KB	6.6 KB	+0.2 KB
other	5.9 KB	5.8 KB	-0.1 KB
everything else	14.3 KB	14.2 KB	-0.1 KB

function	head	change
pgdog::frontend::client::Client::run::{{closure}}	6.9 KB	+0.3 KB
pgdog::frontend::client::query_engine::query::<impl pgdog::frontend::client::query_engine:	5.9 KB	-0.1 KB
pgdog::frontend::client::query_engine::QueryEngine::handle::{{closure}}	4.6 KB	+0.2 KB
tokio::runtime::task::raw::poll	3.7 KB	+0.2 KB
pgdog::frontend::client::query_engine::query::<impl pgdog::frontend::client::query_engine:	3.1 KB	-0.1 KB

sharding-w2:

crate	base	head	change
pgdog::frontend::router	27.9 KB	28.4 KB	+0.4 KB
pgdog::frontend::client::query_engine	21.6 KB	22.2 KB	+0.6 KB
pgdog::backend::pool	17.8 KB	18.7 KB	+0.9 KB
tokio	17.8 KB	17.9 KB	+0.1 KB
aws-lc (TLS crypto)	10.1 KB	10.4 KB	+0.3 KB
other	10.7 KB	10.1 KB	-0.6 KB
pgdog::frontend (other)	9.7 KB	10.0 KB	+0.3 KB
pgdog::backend::server	7.3 KB	7.4 KB	+0.1 KB
everything else	36.0 KB	35.2 KB	-0.8 KB

function	head	change
pgdog::frontend::client::Client::run::{{closure}}	7.4 KB	+0.2 KB
pgdog::frontend::client::query_engine::query::<impl pgdog::frontend::client::query_engine:	6.0 KB	+0.2 KB
pgdog::frontend::client::query_engine::QueryEngine::handle::{{closure}}	4.9 KB	+0.2 KB
pgdog::frontend::router::parser::query::select::<impl pgdog::frontend::router::parser::que	4.1 KB	+0.2 KB
tokio::runtime::task::raw::poll	3.9 KB	+0.1 KB

lb flamegraph vs e9316d2

Open interactive for hover, zoom and search.

Open interactive.

microbench flamegraph vs e9316d2

Open interactive for hover, zoom and search.

Open interactive.

pooler flamegraph vs e9316d2

Open interactive for hover, zoom and search.

Open interactive.

sharding-w2 flamegraph vs e9316d2

Open interactive for hover, zoom and search.

Open interactive.

sharding-w8 flamegraph vs e9316d2

Open interactive for hover, zoom and search.

Open interactive.

Logs

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

Compared a2b6aa9 with e9316d2 on c7a.metal-48xl (spot).