Are they sharing any of this as open source?
Beckman: At this point it's pretty hard to see it. The software that the community is using, none of it is coming from China. It's hard to find, in some sense, on the Web. If you look for some of the pieces like the Kylin ( Linux) operating system, it's not easy to find a community of people where this is being used or shared. It's certainly not prevalent yet. Maybe that's to come. I don't want, in some sense, to sell them short. It's very hard to document code in English if you're really writing essentially in Chinese. There may be language issues preventing them from doing this.
The new Chinese system will use 24 MW (megawatts) at peak when cooling is considered. What are your observations about its power use?
Beckman: That's an awful lot. The raw number is staggering when you think it's about $1 million per year per megawatt. That machine at peak would run $24 million a year in electricity. The goal for exascale is in the 20-30 MW range. In some sense, this shows that if we do nothing, we're stuck at this power rating.
Is there any agreement about how to lower power?
Beckman: There are several promising venues. One is the integration of memory on the chip. Right now, memory accounts for a healthy fraction of that power, and having it external to the CPU wastes power. Pulling it on to the CPU, that memory, with 3D chip stacking or other techniques, will make a big dent.
The other promising technology: Right now, all of our system memory is RAM, and RAM is very inefficient in terms of power. There are technologies that several companies are developing that could use NVRAM. It might not be quite as fast as RAM but the power difference is spectacular, so with that in mind, you can imagine developing systems in the future where some fraction of the memory is actually NVRAM, a smaller fraction of overall memory is RAM, and we get a big power savings. But the thing that we haven't tapped into at all really is managing power as a resource from the software. We just don't have a way right now to automatically move up or down the power in order to take advantage of processors being idle or not idle in a large HPC computation. So there are a lot of software changes that have to happen.
How will the power software management work?
Beckman: Google just wrote a paper, The Tail at Scale. When you do a Google search, it is searching several different servers for little bits of information that are then all pulled together, and that result is then sent back to you. So let's say that there are 20 machines that have to be touched, and a little bit of data from each of the search pieces is assembled and sent back to you. If one of those machines, and this is the part about the tail, comes back with an answer in a slightly longer time, the end result of the query is as long as the longest component. That's frustrating. We find that a little bit interesting, because [Google has] rediscovered what in high-performance computing we have known for a couple of decades, which is this concept of bulk synchronous computation, where you send out hundreds of thousands of tiny work objects to be done, one on each CPU, and if any one of those hundreds of thousands of chips runs slower, any one of them, then your result is as slow as the slowest one.