The definitive Windows SMP method?

Labatyd · 29 Jan 2008 at 09:48

Don't flame me for asking this, as I'm sure the answer to this is nested in an existing thread somewhere - I did search but I found the answers somewhat vague.

With all this negative talk regarding the WinSMP beta client, wouldn't it just be better to run x (where x is the amount of SMP threads you have) instances of the uni-processor client?

I can see the downsides; such as higher memory use, more disk thrashing and more bandwidth use. Just wondering if they still outweigh the PPD hit you take from the beta WinSMP client.

Yeggstry · 29 Jan 2008 at 10:22

From feedback about the affinity changer (see this thread) the talk about the WinSMP is a bit more positive. I'm sure I saw someone running 5 normal clients, perhaps its worth asking them about their PPD

Of course, up to 25% more PPD can be got from using Linux under VMWare...

booyaka · 29 Jan 2008 at 11:10

it appears to be either hit or miss with the win smp client - i've had no problems with it (bar a couple of lost WU's) for months - currently running 2 win smp client with the affinity changer and now pushing 4kppd.

Labatyd · 29 Jan 2008 at 14:09

So am I right in assuming that no one has done a performance/PPD comparision between the SMP and multiple instances on the non-SMP clients?

shadowscotland · 29 Jan 2008 at 14:42

on a single machine I don't think so. (figs below are all Q6600's)

Cob did do some major testing with smp variations - but 4x LinuxSMP (5k ppd) is still out proforming 2x WinSMP (4.1k ppd inc affinity changer) by about 225ppd per core at 3.6Ghz.

At 4.2Ghz Pneumonic has been getting over 5k on very good standard wu's.
but its normally run 3.4 with 4 clients for around 4.2k ppd

normally standard units don't give such high ppd values but there's been a large batch of 3906/7 and as such many of our members have seen a production boost off their non-smp'ing machines

Myself I run a [email protected] for 2.2k with 1x linuxSMP

A.N.Other · 29 Jan 2008 at 17:10

I've always got a far higher ppd crunching winSMP than I have from the equivalent number of standard clients. winSMP runs fine for me under XP (with or without the affinity changer) and I got it working on Vista too, although the installation is a bit more of a fag.

Snapshot · 29 Jan 2008 at 22:48

Labatyd said:
So am I right in assuming that no one has done a performance/PPD comparision between the SMP and multiple instances on the non-SMP clients?

The problem is that there's such a wide variety of console WUs. The very best (Double Gromacs) can rival WinSMP - running four of them on a Quad gives almost the same PPD as two WinSMP WUs with the affinity changer - but there's no guarantee that they're what you'll get. The slowest WUs (SimT Gromacs) may only give a quarter of that. For consistent output, the SMP WUs are the way to go.

jaric · 29 Jan 2008 at 23:15

I've been playing with different clients the last few days and Linux SMP is still king for me.

Q6600 @ 3.7
4 x Linux SMP in 2 virtual machines = 5100ppd
1 x WinSMP = 3200ppd
2 x WinSMP = 4300ppd (ppd the same with/without the affinity changer)
4 x standard clients = 1400ppd

As mentioned previously, ppd can vary on all the clients depending on the work units being crunched, but in general the additional ppd of the SMP clients more than makes up for the extra time spent setting them up and the extra resources they use.

I've just started testing the different clients on my E6750 at work. First WinSMP work unit should be finished this evening. I will give the affinity changer a try tomorrow and then the Linux SMP client in a virtual machine over the weekend.

Yeggstry · 29 Jan 2008 at 23:23

I think the only real way of determining which is best is to do a month average for your own personal configuration. Either that or keep an eye on the posts of people on here for what configs they have and what output they are getting.

For instance, I have been looking at jaric and he is running 4 Linux SMP via VMWare with a Q6600 @ 3.75 and has been consistently been #1 of the single machine folding league since I joined. Granted its probably harder to setup, but then again you can always ask here if you get stuck - that's certainly what I'm going to do.

It would be worth asking Pneumonic how much he is getting a week from his main rig - which from what I remember is a Q6600 @ 3.8 with 5 standard clients. Now, if you can find someone with a similar clock speed and with the WinSMP clients and you will be able to compare similar setups

EDIT: Oh the irony jaric

Pneumonic · 29 Jan 2008 at 23:35

ok well I have finally got the SMP clients running on my main rig after a re-install of XP

both are now 8% in and PPD fluctuates somewhere between 4.8 and 5K according to FahMon.. thats at 3.745

.:MBK:. · 30 Jan 2008 at 20:49

smp gave me much better output overall, however its tendency to drop WUs put me off as im only a part time cruncher so tend to stop and start crunching quite often, combined with the very prompt deadlines its just not something i can crunch at this stage

booyaka · 30 Jan 2008 at 22:04

Pneumonic said:
ok well I have finally got the SMP clients running on my main rig after a re-install of XP both are now 8% in and PPD fluctuates somewhere between 4.8 and 5K according to FahMon.. thats at 3.745

make sure you install the affinity changer as well as it seems to give a good boost!