On Tue, Jul 07, 2009 at 03:28:42PM -0500, Mike Miller wrote:
> In a meeting today we looked at three options from HP and were told that 
> Sun wants to compete with HP and match prices.  The three HP options 
> differed primarily in terms of the CPUs: Intel v. AMD.  With AMD we could 
> get cheaper DDR2 RAM but with Intel we would have to buy pricier DDR3 RAM. 
> All could take SATA 3G HDDs with a RAID controller that would allow us to 
> use RAID1 (simple mirroring) with two 1TB drives per unit.  All were 
> dual-socket quad-core machines, thus 8 cores per unit.  I was thinking 4GB 
> RAM per core or 32GB RAM per unit.  They are 1U rack mountable.
> 
> These are the two machines that we focused on most:
> 
> HP ProLiant DL160 G6 Server (with Intel Xeon X5550)
> http://h10010.www1.hp.com/wwpc/us/en/sm/WF05a/15351-15351-3328412-241644-3328421-3884343.html?jumpid=reg_R1002_USEN
> 
> HP ProLiant DL165 G5 Server (with AMD Opteron 2384)
> http://h10010.www1.hp.com/wwpc/us/en/sm/WF05a/15351-15351-3328412-241644-3328421-3580133.html?jumpid=reg_R1002_USEN
> 
> With the configurations we were shown, the Intel was $2995 and the AMD was 
> $1769, but those prices would increase with added RAM and probably with 
> added HDD space.
> 
> We decided that we need to do some testing to compare the two machines, 
> and we have a couple of machines available.  So I'm going to run some of 
> my genome-scanning code to see how the machines perform.
> 
> 
> Do any of you have any opinions on these options, especially Xeon v. 
> Opteron?  I was told that Xeon was always faster but Opteron was cheaper. 
> For some jobs Xeon was 20% faster and for some it was 100% faster.  My 
> guess is that I will do better, for my work, with a larger number of 
> cheaper processors, but I'm going to do some testing.

One important point that is often overlooked in these Intel-vs-AMD
comparisons is the effective memory bandwidth.  Most genomic
applications are constrained by the memory access time, not as much by
the CPU itself.  Under my desk I have an older Dell dual-core (Xeon
CPU 5140  @ 2.33GHz) and a newer HP quad-core (Xeon E5440  @ 2.83GHz).
My application uses around 1 GB of RAM, and on a theoretical computer
it should scale linearly with the number of cores (it computes some
similarity measure between 10000 'strings'), but in the actual
real-world, the dual core finishes in 8 hours and the quad-core in 7
and a half, despite the fact that all 6 CPUs are 100% busy.  The Dell
dual-core has FB-DIMM 1666 and the HP quad-core has DDR3 1333.  Note
that the quad-core is not the newer i7/Nehalem chips, but the older
'two dual cores on a piece of silicon, fighting over the same FSB'.

Cheers,
florin

-- 
Bruce Schneier expects the Spanish Inquisition.
      http://geekz.co.uk/schneierfacts/fact/163
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
Url : http://mailman.mn-linux.org/pipermail/tclug-list/attachments/20090707/1fa02c06/attachment-0001.pgp