mdk777 wrote:If anyone where to take up your "challenge", they would just prove the existence of the QRB.
I don't think so. I believe Napoleon is looking for
factual, statistical data on how QRB affects the relationship between clients.
I'll bite. But first a few thoughts.
Napoleon wrote:Here comes my challenge:
- Must have -bigadv capable setup. 4, (6), 8, 16, 32 cores et cetera. Real cores? Surreal cores? - Doesn't really matter
- Install the latest beta v7 client
- Set up x amount of uniprocessor slots to match your former bigadv core count
- Report any and all perceived differences - if any. PPD versus FLOPS versus ScientificContribution(TM) and so forth
1) Check. And check.
and both with surreal cores...
2) Why v7? The cores do the work, not the client. I used 6.34+HFM for testing because v7 doesn't calculate the bonus. Monitoring software for real-time tracking of PPD (FahMon, HFM, ect) doesn't work with v7 either.
3) Several thoughts about this.
3a) Hyperthreaded processors (read: virtually all Intel-based -bigadv folders) cannot do so if they follow the recommended practice of 1 classic client per physical core.
SOURCEFor other machines, one can use multiple processors by running multiple clients (one for each real CPU core).
3b) From my experience running 48 classic clients on a 4X 6172 rig would be a waste of resources since all you really need is the average PPD per client X 48 for total PPD. I've run 1, 4, and 8 clients and the variation between them has been minimal. Not to mention one of the 'big dogs' would likely be more amenable to using -smp 40 (w/affinity locked if needed) while running 8 classic clients to get an average TPF for statistical comparison.
3c) The classic WUs need to qualify for QRB.
3d) SMP results would be good too.
4) Not to sound like a broken record but, Flops != Science.
My sole interest (and likely many others) is in PPD, since that is the only measure of scientific value Stanford provides.
Which brings me to comparisons between clients. With the recent release of classic projects that qualify for QRB, I've been able to test this.
So, without further ado...
First machine is a 930 @ 3.8 GHz (4C8T):
UNI 10720 A4 00:27:34 - 4,414.52 PPD (1103.63 X4)
UNI 10721 A4 00:17:45 - 4,637.28 PPD (1159.32 X4)
~~~
SMP 6052 A3 00:03:10 - 16,556.58 PPD
SMP 6067 A3 00:03:17 - 15,682.01 PPD
SMP 7137 A3 00:03:17 - 19,313.38 PPD
~~~
bigadv 2685 A5 00:35:50 - 28,711.51 PPD
bigadv 6900 A5 00:35:55 - 29,611.65 PPD
bigadv 6901 A5 00:36:36 - 27,814.12 PPD
Averaging out the PPD by client type:
Uni @4,525.90 PPD vs SMP @17,183.99 PPD = 3.7968:1
SMP @17,183.99 PPD vs bigadv @ 28,712.43 PPD = 1.6709:1
Second machine is dual 5620 @3.8 GHz (8C16T)
UNI clients same as 930, so 1,131.48 PPD/client X 8 = 9051.84 PPD
~~~
SMP 6052 A3 00:01:36 - 46,099.28 PPD
SMP 6067 A3 00:01:37 - 45,388.24 PPD
SMP 7136 A3 00:01:36 - 56,774.18 PPD
~~~
bigadv 2685 A5 00:18:30 - 77,397.93 PPD
bigadv 6900 A5 00:17:48 - 82,008.13 PPD
bigadv 6901 A5 00:17:49 - 81,893.09 PPD
Averaging out the PPD by client type:
Uni @9,051.84 PPD vs SMP @ 49,420.57 PPD = 5.4597:1
SMP @ 49,420.57PPD vs bigadv @ 80,433.05 PPD = 1.6275:1