The best clouds are genuinely competitive with do-it-yourself high performance computing – and Microsoft's top-tier Azure is the best of the lot.
That's the conclusion from research conducted by Mohammad Mohammadi and Timur Bazhirov of Exabyte and offered as a pre-print at Arxiv.
The two boffins took a simple approach to creating the benchmarks: they ran up the High Performance Linpack on cloud instances, read the dials, and compared that to number 60 on the Top 500 list (at the time of the test), the National Energy Research Science Corporation's Edison unit.
Here are the details of the setups the pair tested (three AWS setups were included for hyper-threaded, non-hyper-threaded, and non-hyper-threaded with placement group configurations):
|Provider||Nodes||Cores||Frequency (GHz)||RAM (GB)||Net (Gbps)|
The AWS configurations generally scaled poorly in terms of speedup obtained by increasing the number of nodes. Non-hyper-threaded machines with placement groups achieved 17.18 times the performance at 32 nodes, compared to one node, which compared to hyper-threaded that achieved 16.41 times the performance at 32 nodes compared to one node.
Of the three Azure configurations tested, the H-series performed best, with 32 nodes delivering 28.33 the performance of a single node; this was, the researchers noted, the best scale-up of all the cloud computing offerings they tested.
Rackspace's Computel-60 instances scaled to 18.55 times the performance at 32 nodes compared to one node; while scaling IBM's SoftLayer from one node to 32 nodes only achieved 4.33 times the speed.
The Edison reference system scaled 27.17 times the speed from one to 32 nodes.
The researchers believe the biggest determinant of scalability is the interconnect in place: “Microsoft Azure outperforms other cloud providers because of the low latency interconnect network that facilitates efficient scaling. SoftLayer has the least favourable speedup ratio at scale, likely because of the interconnect network again.”
Here's their criticism of SoftLayer in detail: “The network quickly saturates at scale, demonstrating the worst performance out of all cases studied. The processor clock speeds are also inferior when compared to other cloud options.”
For most users, of course, scaling to 32 cores is a purely hypothetical exercise, so it's interesting to note outcomes for four nodes, which interest more Register readers.
|AWS||Non-hyperthreading, placement groups||3.15|