Re: [PATCH v6 09/13] zram: split zcomp into separate R/W streams
From: Hillf Danton
Date: Sat Oct 24 2026 - 06:14:41 EST
On Sat, 24 Oct 2026 14:48:32 +0900 Sergey Senozhatsky wrote:
> On (26/10/23 19:40), Hillf Danton wrote:
> [..]
> > > writers vs readers 3 runs
> > >
> > > BASE+PE IOPS 1,271k avg 5.57us sd 1,393us p99.99 27.5us max 1,581,800us
> > > IOPS 1,461k avg 3.43us sd 78us p99.99 23.9us max 227,015us
> > > IOPS 1,444k avg 3.48us sd 222us p99.99 23.9us max 592,628us
> > > PATCHED+PE IOPS 1,457k avg 3.32us sd 18us p99.99 27.5us max 5,620us
> > > IOPS 1,464k avg 3.31us sd 17us p99.99 24.2us max 5,621us
> > > IOPS 1,475k avg 3.29us sd 16us p99.99 26.5us max 5,620us
>
> BASE+PE looked noisy, so I sat down and tested properly (also tested
> zstd and various compression levels).
>
Thanks again for sharing data.
> (PE enabled for all tests)
>
> ######## zstd compression level 3 W-vs-R
> BASE IOPS 1,154k avg 7.84us sd 444us max 697,634us
> IOPS 881k avg 14.16us sd 2,242us max 1,387,000us
> IOPS 704k avg 20.43us sd 3,499us max 2,058,300us
> IOPS 563k avg 29.02us sd 3,910us max 1,649,800us
> IOPS 626k avg 24.79us sd 4,013us max 2,085,800us
> IOPS 976k avg 11.54us sd 1,518us max 1,165,800us
> IOPS 1,178k avg 7.44us sd 22us max 3,103us
> IOPS 1,120k avg 7.39us sd 21us max 7,222us
> IOPS 522k avg 32.51us sd 4,311us max 1,598,300us
> IOPS 581k avg 27.54us sd 3,626us max 1,448,600us
> IOPS 1,160k avg 7.69us sd 352us max 702,099us
> IOPS 1,169k avg 7.52us sd 129us max 281,428us
> IOPS 954k avg 12.06us sd 1,705us max 1,545,600us
> IOPS 1,029k avg 10.24us sd 1,156us max 976,065us
> IOPS 1,169k avg 7.48us sd 24us max 4,305us
> IOPS 1,177k avg 7.45us sd 53us max 118,777us
> IOPS 579k avg 27.43us sd 3,661us max 1,615,800us
> IOPS 1,169k avg 7.52us sd 115us max 228,074us
> IOPS 1,110k avg 8.59us sd 985us max 1,288,000us
> IOPS 1,177k avg 7.45us sd 29us max 72,245us
> IOPS 580k avg 27.74us sd 3,755us max 1,507,000us
> PATCHED IOPS 1,179k avg 7.40us sd 23us max 7,404us
> IOPS 1,179k avg 7.41us sd 22us max 2,857us
> IOPS 1,176k avg 7.40us sd 22us max 6,615us
> IOPS 1,175k avg 7.44us sd 23us max 5,623us
> IOPS 1,176k avg 7.42us sd 24us max 5,625us
> IOPS 1,173k avg 7.42us sd 24us max 5,632us
> IOPS 1,172k avg 7.41us sd 23us max 6,223us
> IOPS 1,173k avg 7.40us sd 25us max 5,848us
> IOPS 1,173k avg 7.40us sd 24us max 7,563us
> IOPS 1,172k avg 7.42us sd 24us max 5,625us
> IOPS 1,175k avg 7.42us sd 23us max 5,626us
> IOPS 1,177k avg 7.42us sd 23us max 4,243us
> IOPS 1,170k avg 7.46us sd 25us max 5,622us
> IOPS 1,178k avg 7.40us sd 22us max 3,213us
> IOPS 1,167k avg 7.43us sd 25us max 7,593us
> IOPS 1,174k avg 7.44us sd 23us max 6,971us
> IOPS 1,179k avg 7.38us sd 23us max 3,477us
> IOPS 1,180k avg 7.39us sd 23us max 2,894us
> IOPS 1,145k avg 7.37us sd 21us max 4,048us
> IOPS 1,177k avg 7.38us sd 22us max 5,620us
> IOPS 1,177k avg 7.40us sd 23us max 5,623us
>
> IOPS: BASE 932k (+-260k), PATCHED 1,174k (+-7k), 21 runs each.
> significant difference at 95% confidence: PATCHED is 241k (25.9%) higher; 95% CI [+136k, +353k], Welch's t-test p = 0.000
> max latency: BASE 3.1..2,085.8 ms, PATCHED 2.9..7.6 ms
> runs with max over 100 ms: BASE 17 of 21, PATCHED 0 of 21; over 1 s: BASE 11 of 21, PATCHED 0 of 21
>
> ######## zstd compression level 12 W-vs-R
> BASE IOPS 690k avg 21.53us sd 3,088us max 1,635,500us
> IOPS 478k avg 36.51us sd 4,385us max 1,941,300us
> IOPS 822k avg 16.13us sd 2,865us max 2,263,800us
> IOPS 1,161k avg 7.73us sd 114us max 190,307us
> IOPS 780k avg 17.56us sd 2,717us max 1,550,500us
> IOPS 1,166k avg 7.66us sd 40us max 94,522us
> IOPS 585k avg 27.51us sd 3,847us max 1,550,600us
> IOPS 1,147k avg 7.97us sd 344us max 496,483us
> IOPS 1,166k avg 7.64us sd 49us max 115,243us
> IOPS 1,163k avg 7.66us sd 23us max 7,006us
> IOPS 1,071k avg 9.42us sd 1,489us max 1,915,500us
> IOPS 962k avg 11.88us sd 2,168us max 1,904,500us
> IOPS 607k avg 26.05us sd 3,538us max 1,591,100us
> IOPS 1,167k avg 7.63us sd 23us max 2,903us
> IOPS 1,157k avg 7.78us sd 198us max 418,930us
> IOPS 1,163k avg 7.66us sd 40us max 100,346us
> IOPS 1,162k avg 7.70us sd 32us max 82,602us
> IOPS 488k avg 35.43us sd 4,340us max 1,786,400us
> IOPS 563k avg 28.85us sd 3,989us max 2,209,500us
> IOPS 843k avg 15.43us sd 2,490us max 1,572,100us
> IOPS 604k avg 26.30us sd 3,560us max 1,555,700us
> PATCHED IOPS 1,166k avg 7.60us sd 23us max 5,623us
> IOPS 1,169k avg 7.58us sd 23us max 5,629us
> IOPS 1,157k avg 7.67us sd 27us max 5,639us
> IOPS 1,166k avg 7.61us sd 24us max 5,644us
> IOPS 1,160k avg 7.64us sd 26us max 7,543us
> IOPS 1,167k avg 7.60us sd 24us max 7,514us
> IOPS 1,160k avg 7.61us sd 25us max 6,643us
> IOPS 1,162k avg 7.65us sd 25us max 7,315us
> IOPS 1,161k avg 7.66us sd 24us max 6,237us
> IOPS 1,158k avg 7.62us sd 27us max 7,688us
> IOPS 1,167k avg 7.58us sd 24us max 5,637us
> IOPS 1,165k avg 7.59us sd 24us max 7,098us
> IOPS 948k avg 7.58us sd 23us max 7,808us
> IOPS 1,166k avg 7.60us sd 24us max 5,621us
> IOPS 1,166k avg 7.58us sd 24us max 5,672us
> IOPS 1,162k avg 7.65us sd 25us max 6,824us
> IOPS 1,165k avg 7.59us sd 24us max 7,078us
> IOPS 1,168k avg 7.58us sd 23us max 6,191us
> IOPS 1,170k avg 7.56us sd 22us max 2,851us
> IOPS 1,171k avg 7.55us sd 22us max 2,834us
> IOPS 1,166k avg 7.59us sd 24us max 5,627us
>
> IOPS: BASE 902k (+-269k), PATCHED 1,154k (+-47k), 21 runs each.
> significant difference at 95% confidence: PATCHED is 252k (27.9%) higher; 95% CI [+142k, +367k], Welch's t-test p = 0.000
> max latency: BASE 2.9..2,263.8 ms, PATCHED 2.8..7.8 ms
> runs with max over 100 ms: BASE 17 of 21, PATCHED 0 of 21; over 1 s: BASE 12 of 21, PATCHED 0 of 21
>
> ######## lz4 W-vs-R
> BASE IOPS 1,422k avg 3.84us sd 418us max 619,495us
> IOPS 1,450k avg 3.57us sd 326us max 609,723us
> IOPS 1,414k avg 3.92us sd 543us max 879,815us
> IOPS 1,451k avg 3.56us sd 253us max 503,428us
> IOPS 1,460k avg 3.40us sd 68us max 135,825us
> IOPS 1,413k avg 3.31us sd 44us max 188,450us
> IOPS 1,432k avg 3.73us sd 503us max 844,228us
> IOPS 1,387k avg 4.20us sd 719us max 961,827us
> IOPS 1,464k avg 3.38us sd 44us max 133,675us
> IOPS 1,469k avg 3.38us sd 22us max 65,812us
> IOPS 1,465k avg 3.40us sd 49us max 111,788us
> IOPS 1,465k avg 3.39us sd 64us max 174,040us
> IOPS 1,467k avg 3.38us sd 36us max 97,821us
> IOPS 1,445k avg 3.63us sd 343us max 671,026us
> IOPS 1,472k avg 3.35us sd 27us max 86,795us
> IOPS 1,473k avg 3.33us sd 44us max 138,096us
> IOPS 1,439k avg 3.71us sd 353us max 619,815us
> IOPS 1,398k avg 3.36us sd 54us max 206,815us
> IOPS 1,471k avg 3.36us sd 15us max 2,832us
> IOPS 1,407k avg 4.02us sd 629us max 1,074,800us
> IOPS 1,465k avg 3.40us sd 63us max 158,779us
> PATCHED IOPS 1,471k avg 3.30us sd 15us max 6,946us
> IOPS 1,473k avg 3.29us sd 16us max 5,618us
> IOPS 1,474k avg 3.29us sd 15us max 5,615us
> IOPS 1,473k avg 3.29us sd 15us max 5,625us
> IOPS 1,440k avg 3.26us sd 13us max 2,829us
> IOPS 1,456k avg 3.30us sd 15us max 5,623us
> IOPS 1,476k avg 3.29us sd 15us max 2,835us
> IOPS 1,464k avg 3.28us sd 16us max 6,063us
> IOPS 1,468k avg 3.32us sd 19us max 6,692us
> IOPS 1,465k avg 3.30us sd 17us max 6,834us
> IOPS 1,468k avg 3.28us sd 16us max 7,268us
> IOPS 1,463k avg 3.31us sd 17us max 6,480us
> IOPS 1,454k avg 3.33us sd 20us max 8,418us
> IOPS 1,452k avg 3.32us sd 19us max 5,621us
> IOPS 1,477k avg 3.27us sd 14us max 2,834us
> IOPS 1,468k avg 3.30us sd 16us max 5,617us
> IOPS 1,475k avg 3.30us sd 16us max 3,884us
> IOPS 1,475k avg 3.30us sd 16us max 5,692us
> IOPS 1,471k avg 3.30us sd 15us max 5,616us
> IOPS 1,473k avg 3.28us sd 16us max 7,014us
> IOPS 1,475k avg 3.28us sd 15us max 5,616us
>
> IOPS: BASE 1,444k (+-27k), PATCHED 1,467k (+-10k), 21 runs each.
> significant difference at 95% confidence: PATCHED is 23k (1.6%) higher; 95% CI [+11k, +35k], Welch's t-test p = 0.001
> max latency: BASE 2.8..1,074.8 ms, PATCHED 2.8..8.4 ms
> runs with max over 100 ms: BASE 17 of 21, PATCHED 0 of 21; over 1 s: BASE 1 of 21, PATCHED 0 of 21
Lets see the data from another pov (feel free to disagree) -- in the android
phone/car cockpit environment, the render frequence is required above 60HZ,
IOW a frame rendered within 16ms, so what share of the 16ms does the IOPS
account for after this patchset, for example when an app in background is
brought foreground?