mersenneforum.org

mersenneforum.org (https://www.mersenneforum.org/index.php)
-   Hardware (https://www.mersenneforum.org/forumdisplay.php?f=9)
-   -   Single vs Dual channel memory (https://www.mersenneforum.org/showthread.php?t=19915)

TObject 2014-12-19 22:43

Single vs Dual channel memory
 
I just run benchmarks with single channel vs dual channel memory. Here they are.
Xeon E3-1225 v3, 3.20 GHz, LL Test on a 36M exponent.

Single channel:
4 workers: 6.5 ms/iter
3 workers: 6.4 ms/iter
2 workers: 6.8 ms/iter
1 worker: 10.0 ms/iter

Dual channel:
4 workers: 3.2 ms/iter
3 workers: 3.5 ms/iter
2 workers: 5.0 ms/iter
1 worker: 9.2 ms/iter

TheMawn 2014-12-20 01:21

This is no surprise. Prime95 is memory bottlenecked so the dual channel giving doubled memory bandwidth is huge.

Interestingly enough, the bottleneck only shows itself with more workers. A single one isn't enough to bottleneck you.

Note that the throughput is almost EXACTLY double with four workers with double the bandwidth, but even at 3 workers, there's a slight discrepancy. It means that your fourth worker, at those speeds, isn't actually giving you much more throughput.

LaurV 2014-12-20 06:53

[QUOTE=TObject;390493]
workers[/QUOTE]
decreasing the time with increasing the number of workers sounds odd, unless you confuse the workers with the threads (single worker, working the same number with many cores).

TObject 2014-12-20 20:14

[QUOTE=LaurV;390558]decreasing the time with increasing the number of workers sounds odd, unless you confuse the workers with the threads (single worker, working the same number with many cores).[/QUOTE]

I am sorry for using wrong terminology, you are right. Instead of "workers" please read "threads."

Thank you.

TObject 2014-12-23 23:18

Another interesting metric.
 
The numbers above were taken running [i]mprime[/i] under [i]Ubuntu 14.04[/i] bare metal.
I decided to measure how much I would lose by virtualizing [i]Ubuntu[/i] under [i]ESXi 5.5[/i]. To my surprise–not that much: with no other virtual machines running, I am only losing about 0.05 of a millisecond per iteration.

I am impressed. I was expecting a much larger penalty.

LaurV 2014-12-24 05:58

Interesting, virtualizing windows with Oracle/Sun's VBox and running P95 in the guest windoze (of course, the host system doing nothing) loses over 2-3% of performance! (of course, in practice this will never be done, but for the sake of metrics).


All times are UTC. The time now is 23:22.

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2021, Jelsoft Enterprises Ltd.