If you listen to IDC talk about HPC, they will talk about “the missing middle”, which is basically a marketing term for a market segment that isn’t being well addressed by HPC vendors with clusters.
It is being addressed by some (hint: the day job) in a variety of ways. An article at InsideHPC by Rich Brueckner gave it a good contextual background, in terms of historical trends in HPC tending to favor the lower cost deployments of processing power. After all, to paraphrase Doug Eadline, a cycle is a cycle is a cycle …
What Intel seems to have discovered, is something we’ve been delivering for years, actually in an expanded sense to what they are piloting. Build muscular workstations with identical spec’s to the servers and add graphics/keyboard/mouse, place them on engineers desks, or right near them. Then let them run.
Of course you want to run Linux on these. If you are like me, you run windows … in a window. See below for my desktop snapshot of exactly this case.
Then you can run your job scheduler, and your cluster code(s). If you tie workgroups of these together with a small IB switch, also located on the floor in the cubicles, you can get a great deal of computing power co-located with the people whom are using it.
Moreover, if you extend this concept a little, this is where Cray’s CX1 fits. Its basically a desk side work-group cluster. The economics of this make sense for those who don’t want or can’t afford the dedicated space in a machine room, or simply can’t afford the high up-front costs of even reasonable sized clusters.
Moreover, we are also starting to see smaller companies forgo purchasing HPC systems, preferring to run on pre-setup remote systems (CRL’s Eka, Sabalcore’s machines, Amazon’s machines). Others are doing cloud type variants, including bare metal clouds from Newservers and others.
I’ve been predicting that the “missing middle” will basically be a bifurcation in the market. That we will be seeing muscular desktops take over the lower end of the cluster market. When you can pack 32-48 processor cores, 1+ GB/s of IO, 1+ GB/s of network IO, and fast graphics along with up to 512 GB ram, and have it sit right next to you … as we can build them … pulling wall current … why would you want a 100 core cluster, which is more costly and painful to run and maintain?
But more to Intel’s point, you can do a similar thing with 12 core units. Bind them together in a deskside cluster. Run a job scheduler.
We’ve been doing this for years. We have multiple customers with these deskside clusters. Build the unit with enough processing power, so it can work on jobs, while doing other day to day work. This is why it should run Linux, and then windows, in a window … you have access to the computational and desktop resources, and can present it in a seamless manner. Not to mention all the benefits of virtualization of the desktop, not the least of which is much better control over virus and infection paths … you can deny most packets coming into or going out of such systems, at very low cost. Moreover, you can recover, trivially, from virus infections, by reverting to a known good VM. Just never store your data on the VM, always on a remote system that has an active backup/snapshot capability. Then you can run whatever you want, and if you get an infection, it isn’t the end of the world, as it is now for so many.
All this said, I do expect desktop/deskside clusters to grow in use. And I expect APU (acceleration processing units) use to continue to grow as accelerators continue to push performance limits.
This is a marvelous heterogenous computing world we are growing into, and while multicore is complex to get right, this non-symmetric world has the right mix of pricing and performance to grow the market. It has to destroy something though to take over its market share. As Rich pointed out, the attack of the killer micros is a repeating paradigm in HPC. It is in the process of repeating again. His analysis is cogent, and I expect that others are going to be agreeing with him quite soon (if not already).
5 years from now, SC15 is going to be quite interested.
Viewed 8256 times by 2003 viewers