I recommend benchmarking/testing
1) you can focus on a significant workload driver on your application and
measure the effect on response time (end to end and per execution)
2) microbenchmark the memory itself - check out intel memory latency
checker
https://software.intel.com/en-us/articles/intelr-memory-latency-checker#whatdoesitmeasure
You can probably try that on one of the nodes first (analyze and correlate
the numbers) then if everything is okay proceed with the rest.
This reminds me of the Exadata X2 days where going for the memory expansion
kit drops the memory frequency to 800 MHz from 1333 MHz. The expansion kits
are usually done on dev/test/qa environments so we could squeeze more
databases on the cluster vs adding more nodes (pricier).
-Karl
On Fri, Mar 31, 2017 at 2:54 PM, Hameed, Amir <Amir.Hameed@xxxxxxxxx> wrote:
We have a four-node ESX farm/cluster where each node is loaded with 512G
of physical memory. These ESX nodes are Cisco UCS blades. We are looking
into adding more memory to each of these ESX hosts to increase the
footprint to 750G/node. One of the data center architects had informed us
last year that per Cisco, the sweet-spot of memory footprint on these
blades is 512GB, after that the memory access speed would drop from 1600MHZ
to 1066MHZ (a drop of ~ 30%). Because all of our VMs host application tiers
and they run a lot of JVM processes, we are concerned that a drop of 30%
will negatively impact our applications’ performance. However, our data
center has been telling us that the drop will not be noticed by the
application which I find hard to believe. We also do not want to add
another ESX host because of licensing cost.
I was wondering if anyone is running their ESX hosts with memory over
512G/host and if they observed any issue with memory speed.
Thank you,
Amir
