On 7/9/2013 8:53 AM, Keith Wesolowski wrote:
Get the documents from Intel describing expected memory bandwidth and
optimal configurations. There are significant -- nay, dramatic --
differences depending on how you populate your controllers and channels.
It's not surprising that having only one CPU would significantly degrade
performance, since the MCs are part of the CPUs. There are also
population rules based on DIMM speed, rank count, power levels, and so
on, and if you do not follow them you will have reduced performance.
Under memtest86+, the *dual* processor board was actually reporting less
bandwidth than the single processor board.
I asked supermicro support what the expected bandwidth would be using
their supported memory, and they were kind enough to run some benchmarks
and send me the results.
They recommended using http://www.cs.virginia.edu/stream/ for benchmarking.
On the X9SRI with 32G (4x8G) they were getting roughly 30GB/s. On the
same board with 48G (6x8G) they were only getting 18GB/s. They said 6
DIMMs isn't balanced, resulting in the lower performance.
On the X9DRi with 64G (8x8G), they were getting roughly 40GB/s, pretty
close to the theoretical maximum.
I'm going to try the same benchmark on my equipment. I originally had
32G in the X9SRi board, I recently picked up an additional 16G as to
increase the ARC, I didn't realize that would result in decreased memory
performance. The motherboard manual doesn't mention balancing the DIMMs,
only which order to install them in depending on how many you have. I
don't really want to add more memory to that box, so I guess I'll have
to decide between the trade-off of increased ARC vs decreased memory
bandwidth.
I'm going to have 64G in the X9DRi before before I deploy it,
unfortunately one of the kits was defective and right now it only has
48G pending RMA.
_______________________________________________
OmniOS-discuss mailing list
OmniOS-discuss@lists.omniti.com
http://lists.omniti.com/mailman/listinfo/omnios-discuss
