Questions about Stanford Client Stats Page

[GreyWhiskers]

I was just looking at the Stanford Client Stats page today, and some of the stats don't pass the "smell test". I added a couple of new columns (Ave # Cores per CPU and Ave x86 GFlops per core). It's obvious why Stanford is trying to wring the most possible out of the Fermi Class GPUs (and the equivalent high-end ATI/AMD GPUs).

Any insights??

- Is the average number of Mac OS X cores per CPU really almost 10?
- For Linux, 8.73 makes sense, since most of the multi-CPU server-class machines run Linux.
- Given that, why is the GFLOPS per Core so low for Mac and Linux, compared with the Windows.
- And, over a population of 300,000 Windows CPUs, I find an average of 2.47 GFlops per core reasonable. See http://www.intel.com/support/processors ... 017346.htm for Intel's info on GFLOPS for US government export control purposes.

Code: Select all

    OS Type      Native TFLOPS*  x86 TFLOPS* Active CPUs Active Cores  Ave#Cores per CPU  AveX86 Gflops per core Total CPUs
        Windows           1,305        1,305     300,028      527,569               1.76                  2.47  4,794,906
       Mac OS X              16           16       5,862       56,395               9.62                  0.28     27,412
          Linux              15           15       4,759       41,553               8.73                  0.36     40,915
        ATI GPU             691        1,458       4,864        4,864               1.00                299.75    360,242
     NVIDIA GPU             710        1,498       3,759        3,759               1.00                398.51    290,320
NVIDIA Fermi GPU          4,670        9,854      13,775       13,775               1.00                715.35    302,968
          Total           7,407       14,146     333,047      647,915                                           5,816,763

[bruce]

Your "smell test" is reasonable, but I think you're assuming you know what data actually collects and how it's interpreted. I don't think that Stanford know as much as it seems and you're misinterpreting what you see. In that chart, I'm pretty sure that a "CPU" is synonymous with a V6 client or a V7 slot, not with actual hardware count. [If it runs a WU, it's considered a CPU.] Moreover, there's a factor of "Hours per day" that you're not taking into account.

Suppose I have a system containing CPUs with 16 cores and I run one CPU:16 client 6 hours a day, those stats will say I contribute 0.25 CPUs. Now if I'm not interested in QRB points, and I'm interested in distorting those figures, I could reconfigure that hardware to run 16 slots, each with a CPU:1 slot/client and run it 24x7 and now that same hardware has 16.0 CPUs. I'll be producing more GFLOPS total and less points using the same number of actual CPUs and actual CPU-cores.

My GUESS is that a lot of Macs run less than 24x7 compared to other platforms. That will decrease the CPU count and probably won't decrease the core count. It should also be noted that the bigadv hardware is (probably) all on Linux and (probably) all runs 24x7, increasing the number of average cores per slot.

That chart needs better footnotes.

[7im]

All recent Macs are all SMP clients, no CPU clients, so the count there is always going to be a little higher, although I didn't expect it to go above 8 cores, as most macs don't have more than that. Where as Windows has both CPU and SMP clients, and a lot of P4s and dual cores still folding. Same goes for linux. While linux does have -bigadv, it also has lots of older hardware still running like Windows.