Are Xeons so bad for GIMPS?

[ric]

I've installed a few days ago a brand new dual Xeon box (Lindenhurst board 2x3.0 Ghz w/HT, 800 Mhz FSB bus, plenty of RAM) with Win2003 Server as OS - and I've to shamefully report that my times per iteration (using v 23.8) are 20% slower than those reported in the benchmark page of the Gimps site. With basically no other process eating up CPU's (surely not at night times).

Processor affinity is set properly, and the two instances running simultaneously show the same (questionable) performance.

Again, turning off via Bios the "logical processor" option (=switching off HT) gives no practical advantage on this machine, since it only achieves a minimal 2/3% improvement (as opposed to the 15/20% mentioned in some thread here), so I'm looking for clues for improvement... or for the definitive sentence giving me the peace of mind of leaving it crunching the way it does.

Hence my question: is Prime95 so optimized for "other" P4's? Someone else experimenting the same behaviour?

TIA,
ric

PS: no, dumping it is *not* an option

[Jeff Gilchrist]

Try stopping one of the Prime95 instances and I bet you will see the iteration time drop quite a bit. My Xeon does the same thing. Xeons have crappy memory buses compared to Opteron systems. My guess is the bus is getting saturated much too easily.

[lpmurray]

I get lousy times on my dell 1600sc duel xeons .081ms with 2 running 10million digit numbers but it drops to .068 with one running. they run on a 533 bus. I bought a asus pc-deluxe board with duel channel memory which I hope will bring my times more in line with the .068 and below plus the board checks components and automatically overclocks up to 15%. When I talked with someone a few years back on the forum they had great times compared to me with the same 2.8 xeon's and came to the conclusion it was the duel channel memory. if your board supports duel channel memory make sure its being utilized.

[Mystwalker]

Maybe setting CPU affinity improves performance?

[Jeff Gilchrist]

Originally Posted by Mystwalker

Maybe setting CPU affinity improves performance?

You can see in ric's post that affinity is set for each client. I am also doing the same.

[Mystwalker]

Oops, overread it...

[geoff]

Some people reported that the new Prime95 version 24.6 is better for dual xeons than 23.8, even though there is no improvement for regular P4s.

[Jeff Gilchrist]

Originally Posted by geoff

Some people reported that the new Prime95 version 24.6 is better for dual xeons than 23.8, even though there is no improvement for regular P4s.

That was definitely the case for my dual Athlon MP system, but sadly not so for my dual Xeons. Using 23.8 and 24.6 produced exactly the same speed.

[Uncwilly]

What if one runs L-L and the other runs TF?

[ric]

To Jeff: indeed, memory seems to be one reason, since stopping one instance gave me some minor improvement: with two instances, I get

Best time for 640K FFT length: 23.196 ms.
Best time for 768K FFT length: 27.920 ms.
Best time for 896K FFT length: 33.521 ms.

while a single one gives me:

Best time for 640K FFT length: 22.257 ms.
Best time for 768K FFT length: 26.789 ms.
Best time for 896K FFT length: 32.211 ms.

However, this is not definitively the kind of solution I'm looking after. Upgrading to 24.6 was one of the proposed activities for the weekend, but since geoff/Jeff's posting, I see no reason for it (furthermore I'm planning to configure NT service version, or 23.5, as a measure towards unattended switchdown/restarts of the box).

Remaining schedule for the weekend includes:
- upgrade win2k3 .inf drivers to the latest version from the Intel site (rationale for it being that w2k3 installation was made from the original disk)
- upgrade the bios (more of a kind of preventive measure, rather than a real need)
- enjoy the coming spring, while the box stands crunching at whatever speed it chooses

I'll keep you posted, if something new and reusable arises. As usual, any other suggestion is welcome.
ric

[TheJudger]

Originally Posted by Jeff Gilchrist

Try stopping one of the Prime95 instances and I bet you will see the iteration time drop quite a bit. My Xeon does the same thing. Xeons have crappy memory buses compared to Opteron systems. My guess is the bus is getting saturated much too easily.

Xeons just have LESS memory buses compared to Opterons...