China's Nebulae supercomputer - zero to second in 3 months
The perils of big-system assembly
HPC Blog Those of you interested in what it really takes to bring up a massive system don't want to miss the "Lessons Learned Deploying the World's First GPU-Based Petaflop System" session.
In it, NVIDIA's senior hardware architect, Dale Southard, discusses his experience with China's Nebulae supercomputer - which, in addition to being the #2 system on the TOP500, was also probably the quickest big build of all time, at around three months total.
Southard, who describes himself as a professional debugger, has quite a track record with big systems. Before NVIDIA, he was at Lawrence Livermore, where he participated in a number of large builds (and probably has the scars to prove it). While the Nebulae super went from bare floor to petabyte processing in record time (about 90 days), it had its share of birthing pains.
In his session, Southard spent some time explaining the differences between small, medium and massive systems. According to him, the 'interesting times' begin in earnest when you move from thousand-node systems to something bigger.
The thousand-node system isn't trivial; it requires considerable - often custom - tooling for management and configuration tasks. But the bigger systems are a whole new world of complexity, mainly due to the fact that things that only rarely (if ever) go wrong on smaller systems malfunction frequently when you amass such a huge array of gear.
Southard related a story about a capacitor that blew up like a mini hand grenade, damaging other components as bits of it wormed their way into nooks and crannies. That's not something you see every day.
He also shared a laundry list of things to check out proactively before you begin big-system assembly. For example: make sure that all the systems have a common BIOS level and that they have the correct processors running at the right speed. When you're talking about such a large number of systems, even stringent quality control can let one or two inconsistent builds slip through. Catching these problems early will save countless hours of troubleshooting down the road.
The videos of the sessions aren't up yet, but I'm told that they should be posted by the end of the week. I'll put in links as soon as I get them... ®
....with most HPC procurements, the vendor doesn't get paid until the machine is accepted and acceptance typically means that it's performing at the promised level. So no money changes hands until all the bugs are ironed out.
GPU is the Way
GPU nodes scale way better faster and cheaper than the puny x86 obsolescence.
or you ask the vendor to benchmark it first
there's no reason why -for a fee- your vendor of 1000 boxes can't rack them up and benchmark them before you take possession of them. That way they get to deal with the BIOS problems and you don't.