dmtcp VS frovedis

I would love to make more use of CRIU. E.g. I considered to use CRIU for my Python preloaded logic (https://github.com/albertz/python-preloaded). Unfortunately, at that point in time, CRIU must be used with root access, which was not an option. However, I see that the PR was merged now, so maybe it works now? (https://github.com/checkpoint-restore/criu/pull/1930)

There is also DMTCP (https://github.com/dmtcp/dmtcp/) but this might have other problems for my use case.

My solution was to use a fork server instead, which works almost equally well. There are not really much downsides with this approach. And this is actually quite simple, and also quite cross-platform (except Windows).

Or maybe DMTCP is a better alternative to CRIU?

All good questions.

1) It is a custom instruction set, you can rean the ISA guide over at https://www.hpc.nec/documentation

2) The main difference in simple terms is that AVX instructions have a fixed vector length (4, 8, 16 etc). With the SX the vector length is flexible so it can be 10, 4, anything up to the max_vlen (up to 256 on the latest ones). Essentially the idea is you have a single instruction that can replace a whole for loop. Without a good compiler though that means you have to re-write your nested loops.

3) There's currently two options when it comes to the compiler, you can use the proprietary NCC or use the open source LLVM fork NEC has. NCC is less compatible than GCC/Clang (particularly modern C++17 is problematic) but has a lot of advanced algorithms for taking your loops and rewriting them and vectorizing them automatically. The LLVM-fork currently supports assembly instruction intrinsics but they are still working on contributing better loop auto-vectorization into LLVM.

4) Porting software is not terribly difficult to get working, but quite a bit harder to get performing very well depending on the type of workload. Since the Scalar core is pretty standard, you can almost always take regular CPU code and get it running (unlike GPU code in general). If you don't leverage the vector processor though, the performance you get will be nothing special, especially at 1.6GHz. Most of the software made for it starts off as being CPU code and is then modified with pragmas or some refactoring to get it running with good performance on the VE. In almost all cases the resulting code still runs on a CPU just fine. One example of a project that supports both in a single code-base is the Frovedis framework[1].

I think the chip deserves a little more interest than it does. It's one of the few accelerators that you can 1) Buy today, right now 2) Has open source drivers [2] 3) Can run tensorflow [3]. The lack of fp16 support really hurt it for Deep Learning but it's like having a 1080 with 48 GB of RAM, still lots of interesting things you can do with that.

[1]: https://github.com/frovedis/frovedis