Sean,
>
> I've seen this trend of poor scaling on the motherboard using three
> different codes. So its not a code-specific problem Also, as Roger
> Kylin mentioned, the Xeons take a bigger hit than the G5.
>
> Do the Opteron boards have dual memory controllers?
Yes the Opteron design is different. Each memory controller is attached
separately with the CPU, so it scales, because there is no contention at any
point, say like the single FSBs in the Xeons or the single memory controller
in the dual G5 towers. But because the Opteron's have a NUMA like
architecture, the shared memory latency is higher than say the Xeon's or the
G5s. But I did get excellent memory B/W for single proc on the G5 about 3.2
GB/s for loads or about 80%. So if you plan to work on single CPU nodes you
should eventually get very good performance from the G5 towers.
> You ask the important question about potential improvements in the
> XServe. Unfortunately Apple is not selling a single CPU Xserve Cluster
> Node (at a decent price).
Apple has to fix the point of contention or the single memory controller
design aspect for the dual nodes and hopefully they have already done so for
Xserve's.
>
> I'll be speaking to the Opteron vendor today and get as much relevant
> information as possible.
The main limiting factor there is the memory speed because their
hypertransport b/w is faster. Also the shared memory latency will only get
worse if they go 4-way or higher nodes. Its still ok for two-way nodes.
-- Avi
>
> On Jan 22, 2004, at 9:09 PM, Avi Purkayastha wrote:
>
>> But its also true, assuming Sean's single processor performance is
>> optimal
>> or close to it, he would not see much scaling anyway, given that the
>> G5 2p
>> system does not scale well, as a single memory controller is shared
>> between
>> the two DIMMs? At least when I did the STREAM benchmark's that was
>> clearly
>> the case. If his code is optimal he should see good b/w for a single
>> proc
>> (3-4 GB/s range) and in the same nbhd for the 2p case.
>>
>> The important question is, has the design been tweaked for improved
>> scaling
>> in a dual Xserve node.
>>
>> -- Avi
>>
>> On 1/22/04 7:19 PM, "Yusuf Abdulghani" <email@hidden> wrote:
>>
>>> Sean,
>>>
>>> Did you profile your code and see where are you spending your time in
>>> the code? What is your memory access pattern? What compiler/compiler
>>> options are you using to compile your code?
>>>
>>> Profile using Shark (profiling tool that comes with CHUD performance
>>> tools) and see if you can determine why you are not seeing as much
>>> scaling as you were expecting to see on the dual G5 system.
>>>
>>> Yusuf
>>>
>>> On Jan 22, 2004, at 3:45 PM, Sean C. Garrick wrote:
>>>
>>>> I've been doing some more benchmarking using two homegrown CFD codes,
>>>> this time with Opterons and PPC-970s (G5).
>>>>
>>>> Here is a typical single processor result :
>>>>
>>>> Small memory code (110 MB): Time Opteron / Time G5 = 2.3
>>>> Opteron: 85 seconds
>>>> G5: 36 seconds
>>>>
>>>> Large memory code (450MB) Time Opteron / Time G5 = 1.9
>>>> Opteron: 454 seconds
>>>> G5: 237 seconds
>>>>
>>>> While the G5 is faster, the Opteron seems to have a better memory
>>>> access/bandwidth.
>>>>
>>>> Also, I ran a job one job on the G5 using 1 processor and another
>>>> using 2 processors. The speed-up was only 50%. I did the same thing
>>>> on
>>>> the Opteron and the speed-up was 83%!!!
>>>>
>>>> I'm not sure what type of board the Opteron was built on but this is
>>>> very surprising as Apple was touting its motherboard as having the
>>>> most bandwidth. I think the performance is quite poor and I'm
>>>> wondering its worth getting 2 CPUs per board or just go with one?
>>>>
>>>> Any ideas?
>>>>
>>>> Thanks,
>>>> Sean
>>>> _______________________________________________
>>>> scitech mailing list | email@hidden
>>>> Help/Unsubscribe/Archives:
>>>> http://www.lists.apple.com/mailman/listinfo/scitech
>>>> Do not post admin requests to the list. They will be ignored.
>>> _______________________________________________
>>> scitech mailing list | email@hidden
>>> Help/Unsubscribe/Archives:
>>> http://www.lists.apple.com/mailman/listinfo/scitech
>>> Do not post admin requests to the list. They will be ignored.
>> _______________________________________________
>> scitech mailing list | email@hidden
>> Help/Unsubscribe/Archives:
>> http://www.lists.apple.com/mailman/listinfo/scitech
>> Do not post admin requests to the list. They will be ignored.
_______________________________________________
scitech mailing list | email@hidden
Help/Unsubscribe/Archives: http://www.lists.apple.com/mailman/listinfo/scitech
Do not post admin requests to the list. They will be ignored.
