Hi Ewan,
To be fair, I was only using old kit that we have at the moment that
clocks in at ~10HS06 per core and from an expectation of them being
worse, a factor 2 difference was massive :) I haven't got round to
benchmarking the new kit yet but these are the 2650v2 chips so will have
a lot better performance as you say.
Thanks,
Mark
On 07/04/14 13:13, Ewan MacMahon wrote:
>> -----Original Message-----
>> From: Testbed Support for GridPP member institutes [mailto:TB-
>> [log in to unmask]] On Behalf Of Mark Slater
>>
>> Thanks for all the info! I sort-of assumed that this was going to be the
>> answer in the end (i.e. that HEPSPEC isn't a good indicator of real life
>> job performance) but I was still a bit surprised that the difference was
>> so much and indeed, that there's such a large dependence on CPU frequency
>> alone. You live and learn I suppose :)
>>
> It might actually be interesting to know what the real price/performance
> difference is though, if someone could come up with a good way to test it.
> This brings us back to the same old discussion about having some sort of
> standard VO reference job.
>
> What did you mean by '*massively* better than any server kit', btw - our
> current generation Xeon 2650v2 chips get ~17 HS06 per (real) core, and
> ~11 per hyperthreading unit (i.e. 22 per physical core if you use the
> hyperthreading, IYSWIM), so ~21 per core from the i5 sounds pretty similar.
>
> On (back of envelope) costs, your desktop is 4x21=84HS06 per box, or
> 84HS06/540GBP = 0.15HS06/£.
>
> One of our Viglen quads is 22HS06*(8cores*2sockets*4boards), or 22SH06*64
> or 1408HS06 per ~£12,000, or 0.12HS06/£
>
> Against that you've got to consider the sheer space that a bunch of desktops
> take up, and the power, and the lack of management cards and so forth, but
> to a first pass, they look like they might, best case, work out a little
> ahead.
>
> Ewan
|