Here’s the first take a look at the configurations of the ASC19 Student Cluster Competition groups. The first element to understand is that every one of the hardware is being generously provided with the aid of Inspur, so this is a lot like a delivery inventory car race, in which all and sundry is strolling frequently the same gear.
So, if absolutely everyone has the identical systems available to them, then why is there the sort of huge kind of clusters inside the competition? Creativity, that’s why. Students have spent months reading and checking out diverse cluster options given the choices to be had to them. Some believe that ‘small is stunning’, specifically when accompanied with the aid of a slew of GPUs, even as others had been looking for better stability between CPU, node remember, and GPU accelerators.
It’s thrilling to have a look at the traits. Systems have become bigger; the average node depend at ASC19 is a full node larger than the average in 2018, with the median growing through nodes. Average and median CPUs and CPU cores have also risen sharply, especially because of Intel increasing their middle counts (best paintings, Intel), however also because of the greater variety of nodes within the systems this 12 months.
We’re seeing a massive increase in memory per node, from 201 GB consistent with node in 2018 to a whopping 384 GB in line with the node in 2019. This is due to new massive reminiscence structures from Inspur – splendid process on their part. And test the memory consistent with cluster figures – 1017 GB in 2018 and 2688 GB in 2019, a big boom over the direction of an unmarried 12 months.
What’s thrilling is that the number of accelerators/GPUs in step with cluster has dropped a bit bit. Some competition observers have expected that the scholar cluster opposition race is in general determined by way of how many accelerators you may cram into a machine. But this isn’t always genuine in step with my information. It all depends on the programs.
The packages this 12 months are a blended bag with regards to GPU-centricity (new phrase on my component). Linpack and HPCG are truely impacted by using what number of GPUs you have in a gadget, but these benchmarks aren’t weighted very highly in relation to the general rankings.
Nvidia is touting their PyTorch GPU tasks, and GPUs can likely be very helpful inside the SR undertaking, however in step with the scholars, you don’t actually need an entire brace of GPUs to finish this assignment.
Which brings us to CESM, which isn’t GPU-centric in any respect in line with the scholars. Since this is a prime part of the scoring, college students in a few instances have eschewed GPUs of their clusters in prefer of more CPU. There’s additionally the truth that Nvidia V100s are quite luxurious and in quick deliver – even though Inspur is imparting four V100s to any group this is interested. Good job Inspur.
Now that we realize the configurations, we’re going to transport directly to meeting the scholars via video interview and getting our first actual consequences of the opposition. Stay tuned!