In another post, someone said that Linux is faster than Windows for CP.net crunching.

How much faster?

I know that there are 64-bit Linux available (my main machine is 64-bit) but is there a 64-bit client for CP.net?

If not, how fast is the 32-bit client compared with Boinc 4.19 running on WinXP 2nd edition?

Matthew

> In another post, someone said that Linux is faster than Windows for CP.net
> crunching.
>
> How much faster?
>
> I know that there are 64-bit Linux available (my main machine is 64-bit) but
> is there a 64-bit client for CP.net?
>
> If not, how fast is the 32-bit client compared with Boinc 4.19 running on
> WinXP 2nd edition?
>
IMO, it depends on your processor. For AMD, I've seen little if any speed difference between Linux and Windows, maybe slightly slower in Linux but others may disagree. On Intel, at least on my P4, Linux appears to be 5 to 10% faster than Windows. It's hard to compare accurately unless you run several in Windows and several in Linux since there are significant run to run differences in model speed depending on the parameters given.

There is no 64 bit client yet, and as Dave Frame stated in your other thread, it isn't just around the corner either given other IT resource priorities. I look forward to it however, as, given my experience in sulphur alpha, one could finish about 3 models in the time it presently takes to do 2.

Yes I asked that question. I did simply because I have a linux box P4 3 Gig HT no sse3 and it was working through phase 1 at 0.85 s/TS. 7 Days it took. On a Win 2003 server P4 3 Gig with sse3 (so a faster box) its running at 3.18 s/TS. The first pc needed 7 days to complete the second PC is indciating it needed 29 days to complete. Perhaps the work unit is very different I really don't know but that is quite a speed difference. Especially as the linux box is running apache, dns, smb, firewall etc......whereas the windows box is just boincing. Not sure this helps advance towards a proper answer to the question....but its a little evidence perhaps.
<img src="http://www.boincsynergy.com/images/stats/comb-1091.jpg"></img><br><img src="http://www.iantighe.com/setisig.jpg"></img><img border="0" src="http://boinc.mundayweb.com/one/teamStats.php?userID=1602&amp;prj=1&amp;trans=off">

&gt; Yes I asked that question. I did simply because I have a linux box P4 3 Gig HT
&gt; no sse3 and it was working through phase 1 at 0.85 s/TS. 7 Days it took. On a
&gt; Win 2003 server P4 3 Gig with sse3 (so a faster box) its running at 3.18 s/TS.
&gt; The first pc needed 7 days to complete the second PC is indciating it needed
&gt; 29 days to complete. Perhaps the work unit is very different I really don't
&gt; know but that is quite a speed difference. Especially as the linux box is
&gt; running apache, dns, smb, firewall etc......whereas the windows box is just
&gt; boincing. Not sure this helps advance towards a proper answer to the
&gt; question....but its a little evidence perhaps.
&gt;
It appears as if no optimization is made for SSE3. At least there is little discernable difference between a Prescott and Northwood P4 in CPDN performance. Of course, comparisons of Northwoods and Prescotts indicate some performance improvements of Prescott over Northwood in some apps not utilizing SSE3, and degradation in other apps, so a small difference could be due to random parameters used in the model, or differences in the way CPDN runs on one processor vs. the other (use of cache for example).

However, 0.85 sec/TS is not possible with any PC (maybe a P4 at 5 GHz). What probably happened is that the duration of the run got reset through some problem, so the sec/TS got to be artificially small. I've seen that happen on my Linux P4 and Linux Xeons. The Linux P4 might be running two models in hyperthreading at 2.4 sec/TS apiece. Suddenly some glitch happens and one of them thinks it is running at 1.3 sec/TS (or 0.5 or whatever) because the duration has been reset. In that case, the sec/TS will go up for each trickle as they accumulate.

At 0.85 sec/TS, the entire model would complete in 7 days. At 3.18 the entire model would complete in 29 days. No model has been completed in 7 days. For comparison purposes:

2 sec/TS = 6 hours trickle duration = 6 days phase duration = 18 days per model
3 sec/TS = 9 hours trickle duration = 9 days phase duration = 27 days per model

That's assuming running 24x7 and only CPDN. The very fastest model run would likely come from some highly overclocked P4, HT turned off, maybe 1.2 sec/TS.

I don't think the very fastest would come from a P4 - I'm more inclined to place my bets on a fast AMD processor (such as the AMD64 FX) Those are also much more overclock-friendly, AFAIK.

Excluding high-end servers and supercomputers, of course...

Matthew

Hi,
I've noted a difference on my P4 3ghz (prescott): a model on Windows is about 3.5 S/TS and on Linux, the s/ts is always under 3 (2.91 on my current model). I'm saving about 100 hours of computation on Linux (it's a mean value, as the parameters can slow or accelerate a model)
I have noted the same thing on my old P4 2ghz: s/TS=~2.8 on Linux and ~3.4 on Windows

Ah Ok thanks for the information there. My linux box has now moved onto phase 2 of its model as far as I can tell. Interestingly its doing 49.06 s/TS.....and needs 442 days to complete.......so....I wont trust these numbers perhaps! I use boincview btw so perhaps its not quite accurate either. Thx again.
<img src="http://www.boincsynergy.com/images/stats/comb-1091.jpg"></img><br><img src="http://www.iantighe.com/setisig.jpg"></img><img border="0" src="http://boinc.mundayweb.com/one/teamStats.php?userID=1602&amp;prj=1&amp;trans=off">

&gt; I don't think the very fastest would come from a P4 - I'm more inclined to
&gt; place my bets on a fast AMD processor (such as the AMD64 FX) Those are also
&gt; much more overclock-friendly, AFAIK.
&gt;
Yes, but they use an Intel compiler for CPDN and it optimizes much better for Intel processors than AMD. The fastest AMD PCs are generally in the 1.6 sec/TS range, while the fastest P4s when running only one model, with HT off, are closer to 1.3 sec/TS (in Linux). But it would be interesting to see what an FX-57 would do.

Why do they do that, I wonder? The same reason Dell snubs AMD to this day? They can't honestly say there isn't demand, that they are a small player, etc. they're up to 1/3 market share now I think! Considering who they're up against, they are doing great! They really must have something going for them technology-wise, or everyone would go with the big boy Intel.

You'd think it would be worth their while to release an AMD-enhanced version as well. Isn't there a more "neutral" compiler available than the Intel one they are using?

Matthew

&gt; Why do they do that, I wonder? The same reason Dell snubs AMD to this day?
&gt; They can't honestly say there isn't demand, that they are a small player, etc.
&gt; they're up to 1/3 market share now I think! Considering who they're up
&gt; against, they are doing great! They really must have something going for them
&gt; technology-wise, or everyone would go with the big boy Intel.
&gt;
&gt; You'd think it would be worth their while to release an AMD-enhanced version
&gt; as well. Isn't there a more "neutral" compiler available than the Intel one
&gt; they are using?
&gt;
This goes back a long ways. I think it is down to cost. They had an Intel compiler. Compiling the CPDN code with it there was a slight performance degradation for A64s, about the same performance for Athlons, and a <b>huge</b> performance improvement for Intel over "Classic" (pre-BOINC) CPDN throughput. So for a low cost, they got a big boost in the project. I believe Carl, the former head IT guy on CPDN, tried other compilers. But time for testing, large amount of IT work, and potential difficulties in deployment of apps led to abandonment of those plans. I think later Intel compilers do a better job with AMD processors, so if they recompile with a later compiler, that may be good. But the big, big improvement will come from 64 bits. If you do a message board search for compiler, you will see a few thread and some of the effort that went on.

geophi wrote:
&gt; The fastest AMD PCs are generally in the 1.6
&gt; sec/TS range, while the fastest P4s when running only one model, with HT off,
&gt; are closer to 1.3 sec/TS (in Linux).

I think your estimated fastest speed of the AMD section with 1.6 sec/TS at its best isn't very realistic. I use a moderate overclocked AMD Athlon64 3000+ Winchester currewntly at 2208MhZ with DDR400 Infineon Memory and I got about ~1.9 sec/Ts on average. So I would say already AMD Athlon64 3800+ or 4000+ with its 2.4Ghz should reach not overclocked an processing speed of approx ~1.7 sec/TS. Since one can also ocerclock these processors very well their expected speed should reach 1.5 maybe 1.4 sec/TS. I think its not so unrealistically.

Fourthermore one should keep in mind the much lower power consumption and therefore the much lower temperatures of the AMD section. In my opinion a very important issue towards the efficiency in office use.

Ciao

&gt; I think your estimated fastest speed of the AMD section with 1.6 sec/TS at its
&gt; best isn't very realistic. I use a moderate overclocked AMD Athlon64 3000+
&gt; Winchester currewntly at 2208MhZ with DDR400 Infineon Memory and I got about
&gt; ~1.9 sec/Ts on average. So I would say already AMD Athlon64 3800+ or 4000+
&gt; with its 2.4Ghz should reach not overclocked an processing speed of approx
&gt; ~1.7 sec/TS. Since one can also ocerclock these processors very well their
&gt; expected speed should reach 1.5 maybe 1.4 sec/TS. I think its not so
&gt; unrealistically.
&gt;
&gt; Fourthermore one should keep in mind the much lower power consumption and
&gt; therefore the much lower temperatures of the AMD section. In my opinion a very
&gt; important issue towards the efficiency in office use.
&gt;
My statement about 1.6 sec/TS was just an observation of the fastest A64 I've seen. DaveN's overclocked FX53 does about that. I'm not sure how much he has overclocked it though. No doubt some Paris or Venice core A64 with overclocking headroom could go considerably lower. I just haven't seen the results of that yet.

As for power consumption, no doubt about the A64s (particularly Venice or Paris) being much more efficient than the ridiculous Prescotts, but this thread was about performance in Linux vs. Windows, which has now become Intel vs. AMD. :-)