Are Palomino's as fast as 1 MB Xeons at SETI?!?!

[Sukhoi]

I was just thinking of this. Insane3D's 1.4 GHz Athlon 4 is doing 3+ hour WU's. Now get 2 of these in a dual board and you're doing 16 WU/day from 2800 Mhz! :Q gaijin's quad 1 MB Xeon 700 MHz server was also putting out 16 WU/day from 2800 MHz. That's fast! 😀

 

[Baldy18]

Excellent. /meniacle laugh

 

[Thyme]

The new ones based on the P4?

 

[Sukhoi]

No, I meant PIII Xeons. Currently P4 Xeons are the same speed as regular P4's (though they still rock) because both have 256 KB of L2 cache.

 

[ElFenix]

i think you'll see a reduction in WU/day out of dual pallys since there isn't as much RAM bandwidth for each to run its hardware prefetch on (of course... SETI is probably a very predictable program... wonder how a pally would perform in a KT133a board?). unless of course there is dually nforce. *droooolllls*

 

[BurntKooshie]

It's not just not enough bandwidth to run HW prefetch -- Palomino's suffer the same problem that Intel boards do. You can spout the "point to point" bus all you want, but that is only useful for applications where the two processors are sharing the same data (i.e., they are running a multithreaded applications).

With S@H, it's two clients running, because it's not multithreaded. The dedicated snoop-bus does nothing in this case. While the performance benefit from going dually with P3s is relatively minor (main memory bandiwth contention), with P4's and Palominos, the performance degredation from adding another processesor is less (good thing), due to increased bandwidth.

I'm saying this only because it's not so much the HW prefetch that will cause Palominos performance to be lessened when adding another processor, but simply the lack of bandwidth/low_latencies itself. The issue is the exact same for Pallys that it is for P3s, except Pallys have more bandwidth. Given the soaring clock speeds, eventually the Pally will show return to scales similar to the P3, due to the main_memory:clock-rate ratios being similar.

 

[cnhoff]

From time to time I use a dual PIII-500 with SetiDriver and two clients, and I get wu times of about 12 hours. I think that is not too bad and you can surely expect at at least 3,5 hours/wu from a dual Palomino@1400.

 

[BurntKooshie]

Well, ideally, a dual config would get it in half the time. From what I'm hearing, a Palomino does it in ~3.5 hours. Ideally, the rate of completion would double ,meaning average time per WU for completion in the machine would be cut in half, except from each clients perspective, where it would be the same (because it's not multithreaded).

On the other hand, reality tells us that the rate of completion will be somewhat south of double, so the time for eacch client to finish one WU will increase a little, but overall output goes up.

So the average time in a dual config might go up to something like 4.5 hours for each client, the output will be 2 WUs in 4.5 hours. Realistically, I'd imagine that a dual Pally would get ~2 WU's every 4.5 hours, or average (for the machine, not the individual client) about 1WU every 2.25 hours (ish).