PrimeGrid
Please visit donation page to help the project cover running costs for this month

Toggle Menu

Join PrimeGrid

Returning Participants

Community

Leader Boards

Results

Other

drummers-lowrise

Advanced search

Message boards : Number crunching : Strange timings on AMD 3700X

Author Message
Profile Crun-chiProject donor
Volunteer tester
Avatar
Send message
Joined: 25 Nov 09
Posts: 3114
ID: 50683
Credit: 76,797,694
RAC: 7,338
Eliminated 1 conjecture "k"Found 1 prime in the 2018 Tour de PrimesFound 1 prime in the 2019 Tour de PrimesFound 1 prime in the 2020 Tour de PrimesFound 1 prime in the 2021 Tour de PrimesFound 1 prime in the 2022 Tour de Primes321 LLR Silver: Earned 100,000 credits (229,492)Cullen LLR Silver: Earned 100,000 credits (110,733)PPS LLR Sapphire: Earned 20,000,000 credits (23,633,806)PSP LLR Silver: Earned 100,000 credits (104,385)SoB LLR Silver: Earned 100,000 credits (106,117)SR5 LLR Amethyst: Earned 1,000,000 credits (1,240,750)SGS LLR Amethyst: Earned 1,000,000 credits (1,325,207)TRP LLR Gold: Earned 500,000 credits (626,755)Woodall LLR Silver: Earned 100,000 credits (122,944)321 Sieve (suspended) Silver: Earned 100,000 credits (104,900)Cullen/Woodall Sieve (suspended) Ruby: Earned 2,000,000 credits (2,000,599)Generalized Cullen/Woodall Sieve (suspended) Gold: Earned 500,000 credits (515,556)PPS Sieve Jade: Earned 10,000,000 credits (11,889,685)TRP Sieve (suspended) Silver: Earned 100,000 credits (255,612)AP 26/27 Ruby: Earned 2,000,000 credits (2,583,960)GFN Sapphire: Earned 20,000,000 credits (23,337,083)WW Amethyst: Earned 1,000,000 credits (1,088,000)PSA Turquoise: Earned 5,000,000 credits (7,522,050)
Message 155999 - Posted: 19 Jun 2022 | 22:56:00 UTC

I run all tests under Linux and using mprime. Found strange timings but only on 640K and 8 workers, one cpu core per worker (SMT is disabled, Linux is OS) and found no explanation. Any idea?

Prime95 64-bit version 30.8, RdtscTiming=1
Timings for 640K all-complex FFT length (8 cores, 1 worker): 0.43 ms. Throughput: 2317.71 iter/sec.
Timings for 640K all-complex FFT length (8 cores, 2 workers): 0.54, 0.54 ms. Throughput: 3704.59 iter/sec.
Timings for 640K all-complex FFT length (8 cores, 4 workers): 1.05, 1.04, 1.04, 1.04 ms. Throughput: 3831.54 iter/sec.
Timings for 640K all-complex FFT length (8 cores, 8 workers): 4.22, 4.22, 4.22, 2.59, 4.20, 4.20, 2.38, 4.21 ms. Throughput: 2230.50 iter/sec.

Timings for 640K all-complex FFT length (8 cores, 1 worker): 0.43 ms. Throughput: 2330.67 iter/sec.
Timings for 640K all-complex FFT length (8 cores, 2 workers): 0.58, 0.58 ms. Throughput: 3428.54 iter/sec.
Timings for 640K all-complex FFT length (8 cores, 4 workers): 1.13, 1.13, 1.14, 1.13 ms. Throughput: 3530.72 iter/sec.
Timings for 640K all-complex FFT length (8 cores, 8 workers): 4.20, 4.24, 2.60, 4.08, 2.42, 4.35, 4.37, 4.24 ms. Throughput: 2210.83 iter/sec.

Prime95 64-bit version 29.8, RdtscTiming=1
Timings for 640K all-complex FFT length (8 cores, 1 worker): 0.43 ms. Throughput: 2301.15 iter/sec.
Timings for 640K all-complex FFT length (8 cores, 2 workers): 0.55, 0.55 ms. Throughput: 3626.02 iter/sec.
Timings for 640K all-complex FFT length (8 cores, 4 workers): 1.07, 1.07, 1.06, 1.06 ms. Throughput: 3751.43 iter/sec.
Timings for 640K all-complex FFT length (8 cores, 8 workers): 4.29, 2.77, 4.18, 4.32, 4.26, 3.45, 3.01, 4.34 ms. Throughput: 2152.02 iter/sec.

This machine is dedicated just for crunching and nothing else. Also as you can see in red color, not always same core are faster then rest of them.

____________
92*10^1439761-1 NEAR-REPDIGIT PRIME :) :) :)
4 * 650^498101-1 CRUS PRIME
314187728^131072+1 GENERALIZED FERMAT
Proud member of team Aggie The Pew. Go Aggie!

GDB
Send message
Joined: 15 Nov 11
Posts: 280
ID: 119185
Credit: 3,383,664,451
RAC: 3,840,659
Discovered 2 mega primesFound 1 prime in the 2022 Tour de Primes321 LLR Jade: Earned 10,000,000 credits (13,480,830)Cullen LLR Turquoise: Earned 5,000,000 credits (5,837,403)ESP LLR Turquoise: Earned 5,000,000 credits (5,600,055)Generalized Cullen/Woodall LLR Turquoise: Earned 5,000,000 credits (7,361,683)PPS LLR Jade: Earned 10,000,000 credits (17,608,655)PSP LLR Turquoise: Earned 5,000,000 credits (7,468,468)SoB LLR Sapphire: Earned 20,000,000 credits (21,450,372)SR5 LLR Turquoise: Earned 5,000,000 credits (6,839,538)SGS LLR Turquoise: Earned 5,000,000 credits (5,688,853)TRP LLR Jade: Earned 10,000,000 credits (10,398,149)Woodall LLR Turquoise: Earned 5,000,000 credits (5,099,498)321 Sieve (suspended) Turquoise: Earned 5,000,000 credits (5,834,160)Cullen/Woodall Sieve (suspended) Ruby: Earned 2,000,000 credits (2,647,915)Generalized Cullen/Woodall Sieve (suspended) Jade: Earned 10,000,000 credits (11,373,349)PPS Sieve Double Silver: Earned 200,000,000 credits (498,780,133)Sierpinski (ESP/PSP/SoB) Sieve (suspended) Ruby: Earned 2,000,000 credits (4,080,038)TRP Sieve (suspended) Turquoise: Earned 5,000,000 credits (7,803,337)AP 26/27 Double Silver: Earned 200,000,000 credits (200,108,285)GFN Double Silver: Earned 200,000,000 credits (355,744,232)WW Double Silver: Earned 200,000,000 credits (327,540,000)PSA Double Amethyst: Earned 1,000,000,000 credits (1,863,027,834)
Message 156001 - Posted: 20 Jun 2022 | 2:11:52 UTC

It seems that when you have more than 4 workers, your L3 cache usage has exceeded your 32 MB of L3 cache.
So your 8 workers are causing the memory subsystem to thrash and all memory requests are going to slow memory, instead of fast L3 cache.
____________

Profile Crun-chiProject donor
Volunteer tester
Avatar
Send message
Joined: 25 Nov 09
Posts: 3114
ID: 50683
Credit: 76,797,694
RAC: 7,338
Eliminated 1 conjecture "k"Found 1 prime in the 2018 Tour de PrimesFound 1 prime in the 2019 Tour de PrimesFound 1 prime in the 2020 Tour de PrimesFound 1 prime in the 2021 Tour de PrimesFound 1 prime in the 2022 Tour de Primes321 LLR Silver: Earned 100,000 credits (229,492)Cullen LLR Silver: Earned 100,000 credits (110,733)PPS LLR Sapphire: Earned 20,000,000 credits (23,633,806)PSP LLR Silver: Earned 100,000 credits (104,385)SoB LLR Silver: Earned 100,000 credits (106,117)SR5 LLR Amethyst: Earned 1,000,000 credits (1,240,750)SGS LLR Amethyst: Earned 1,000,000 credits (1,325,207)TRP LLR Gold: Earned 500,000 credits (626,755)Woodall LLR Silver: Earned 100,000 credits (122,944)321 Sieve (suspended) Silver: Earned 100,000 credits (104,900)Cullen/Woodall Sieve (suspended) Ruby: Earned 2,000,000 credits (2,000,599)Generalized Cullen/Woodall Sieve (suspended) Gold: Earned 500,000 credits (515,556)PPS Sieve Jade: Earned 10,000,000 credits (11,889,685)TRP Sieve (suspended) Silver: Earned 100,000 credits (255,612)AP 26/27 Ruby: Earned 2,000,000 credits (2,583,960)GFN Sapphire: Earned 20,000,000 credits (23,337,083)WW Amethyst: Earned 1,000,000 credits (1,088,000)PSA Turquoise: Earned 5,000,000 credits (7,522,050)
Message 156003 - Posted: 20 Jun 2022 | 3:49:29 UTC - in response to Message 156001.

It seems that when you have more than 4 workers, your L3 cache usage has exceeded your 32 MB of L3 cache.
So your 8 workers are causing the memory subsystem to thrash and all memory requests are going to slow memory, instead of fast L3 cache.

I have no problem with that, and I use 4 workers. But if as you say memory subsystem "fail" why then few cores are double faster then others?
____________
92*10^1439761-1 NEAR-REPDIGIT PRIME :) :) :)
4 * 650^498101-1 CRUS PRIME
314187728^131072+1 GENERALIZED FERMAT
Proud member of team Aggie The Pew. Go Aggie!

GDB
Send message
Joined: 15 Nov 11
Posts: 280
ID: 119185
Credit: 3,383,664,451
RAC: 3,840,659
Discovered 2 mega primesFound 1 prime in the 2022 Tour de Primes321 LLR Jade: Earned 10,000,000 credits (13,480,830)Cullen LLR Turquoise: Earned 5,000,000 credits (5,837,403)ESP LLR Turquoise: Earned 5,000,000 credits (5,600,055)Generalized Cullen/Woodall LLR Turquoise: Earned 5,000,000 credits (7,361,683)PPS LLR Jade: Earned 10,000,000 credits (17,608,655)PSP LLR Turquoise: Earned 5,000,000 credits (7,468,468)SoB LLR Sapphire: Earned 20,000,000 credits (21,450,372)SR5 LLR Turquoise: Earned 5,000,000 credits (6,839,538)SGS LLR Turquoise: Earned 5,000,000 credits (5,688,853)TRP LLR Jade: Earned 10,000,000 credits (10,398,149)Woodall LLR Turquoise: Earned 5,000,000 credits (5,099,498)321 Sieve (suspended) Turquoise: Earned 5,000,000 credits (5,834,160)Cullen/Woodall Sieve (suspended) Ruby: Earned 2,000,000 credits (2,647,915)Generalized Cullen/Woodall Sieve (suspended) Jade: Earned 10,000,000 credits (11,373,349)PPS Sieve Double Silver: Earned 200,000,000 credits (498,780,133)Sierpinski (ESP/PSP/SoB) Sieve (suspended) Ruby: Earned 2,000,000 credits (4,080,038)TRP Sieve (suspended) Turquoise: Earned 5,000,000 credits (7,803,337)AP 26/27 Double Silver: Earned 200,000,000 credits (200,108,285)GFN Double Silver: Earned 200,000,000 credits (355,744,232)WW Double Silver: Earned 200,000,000 credits (327,540,000)PSA Double Amethyst: Earned 1,000,000,000 credits (1,863,027,834)
Message 156004 - Posted: 20 Jun 2022 | 4:16:31 UTC

A few of the cores may be faster because they are able to keep all of their data in L3 cache.
Everybody else has to access slow main memory. Life isn't fair. First come, first served.
____________

Post to thread

Message boards : Number crunching : Strange timings on AMD 3700X

[Return to PrimeGrid main page]
DNS Powered by DNSEXIT.COM
Copyright © 2005 - 2022 Rytis Slatkevičius (contact) and PrimeGrid community. Server load 0.40, 0.55, 0.54
Generated 11 Aug 2022 | 11:02:26 UTC