Considerations for choice of building blocks for your SMP
May 13, 2013
If you’ve had any experience in designing solutions for IT deployment (and even more so, specifically with HPC solutions), you’d know that there are always trade-offs. Budget is almost always finite and tight, and you need to make sure that you use the scarce dollars to yield the best for your workloads. If you are lucky, your mix of workloads can be characterized – and you can apply this knowledge to your solution design.
So, you’ve decided to deploy an Intel-based SMP solution, and you are asking yourself “which processor series should I be using?” the answer depends on a few factors. For the sake of this discussion, I will ignore the generation-old Westmere processors (e.g. E7-x8xx) and focus on the latest generations: E5-2600 and E5-4600. Clearly, if you need a higher clock speed, a lower cost per flop, or high density, then the E5-2600 is your choice. On the DIMMs per socket, the two E5-* series can provide as high as 12 DIMMs per socket, so (depending on choice of server model), that’s a tie. There are, however, cases where the E5-4600 provides an advantage: mainly in cases where it is not just the RAM/core ratio that you care about, but also the total RAM per NUMA-node. With 4-socket nodes using the E5-46xx, you could get as high as 1.5TB RAM per physical server (or, when aggregated by vSMP Foundation, per NUMA-node). That might give you an advantage if your typical workload uses a low core count but a significant amount of RAM – e.g. an analytics database.
To summarize, if you can characterize your typical SMP workloads, you could potentially optimize the solution’s building blocks to maximize your overall performance – and doing so might require thinking not only of the GHz and FLOPS, but also on the other aspects of the hardware components of choice such as total RAM per node.
Try us. Check our web-site for more details.