A Beta Tomorrow
June 10, 2012
Posted by on
I originally envisioned a beta release this summer. Well, that’s now. The new date is before next year. It’s a more real date (at least for me) as I am saying it in public at a conference in a few days.
These two talks are interesting.
http://research.microsoft.com/apps/video/dl.aspx?id=165730 “Evolving GPUs into a Substrate for Cloud Computing” – The takeaway is that GPUs are changing. SoC processors with integrated GPUs (e.g. AMD Fusion APU, Intel Core, Nvidia Tegra) may lack the sexy high peak throughput of big discrete cards. However, they may also have much lower penalties for data transfers. The talk showcased an example of porting memcached indexing to an AMD Fusion APU. Yes, it does run significantly faster even with the I/O costs and control flow divergence. A big discrete like an ATI HD 5870 has terrible performance in this use case.
http://www.youtube.com/watch?v=NWSZ4c9yqW8 “Simon Peyton Jones: Data Parallel Haskell ” – The takeaway is pure functional is good for parallel code but bad for locality of data. That’s bad as data movement costs often dominate performance in practice.
The many-core accelerator designs all seem to fail in the marketplace. STI Cell. Larrabee. They are too expensive for what they do, the Swiss Army knife effect of not excelling at anything. What seems to win are the multi-core CPU and the GPU. These are very good at what they do.
It is the bazaar of CPUs and GPUs over the cathedral of many-core.
There’s a lot of interest now in “device fission”. I see this as part of virtualization which is really about two things which are both the same: sell more with less; increase device utilization. It’s like slicing the pizza and selling the slices. You make more money that way and better utilize the pizza too.
I haven’t seen any talk of “device fusion”. Why can’t a cloud look like a big stream processor? If there were virtual fused compute devices, then applications could scale arbitrarily.
But maybe the reason “device fusion” isn’t considered is that, like the nuclear counterpart, it’s something cool that is impractically difficult. That’s a little like the many-core vision of Swiss Army knife homogeneous computing devices.
So the lesson here is that heterogeneity is real. It’s ugly. But that’s diversity in the bazaar. We want cathedrals because they appeal to inner drives to seek beauty through perfection and order. That’s a trap.