Björn Töpel wrote:
On Tue, 28 Jan 2020 at 03:14, Palmer Dabbelt palmerdabbelt@google.com wrote:
There's four patches here, but only one of them actually does anything. The first patch fixes a BPF selftests build failure on my machine and has already been sent to the list separately. The next three are just staged such that there are some patches that avoid changing any functionality pulled out from the whole point of those refactorings, with two cleanups and then the idea.
Maybe this is an odd thing to say in a cover letter, but I'm not actually sure this patch set is a good idea. The issue of extra moves after calls came up as I was reviewing some unrelated performance optimizations to the RISC-V BPF JIT. I figured I'd take a whack at performing the optimization in the context of the arm64 port just to get a breath of fresh air, and I'm not convinced I like the results.
That said, I think I would accept something like this for the RISC-V port because we're already doing a multi-pass optimization for shrinking function addresses so it's not as much extra complexity over there. If we do that we should probably start puling some of this code into the shared BPF compiler, but we're also opening the doors to more complicated BPF JIT optimizations. Given that the BPF JIT appears to have been designed explicitly to be simple/fast as opposed to perform complex optimization, I'm not sure this is a sane way to move forward.
Obviously I can only speak for myself and the RISC-V JIT, but given that we already have opened the door for more advanced translations (branch relaxation e.g.), I think that this makes sense. At the same time we don't want to go all JVM on the JITs. :-P
I'm not against it although if we start to go this route I would want some way to quantify how we are increasing/descreasing load times.
I figured I'd send the patch set out as more of a question than anything else. Specifically:
- How should I go about measuring the performance of these sort of optimizations? I'd like to balance the time it takes to run the JIT with the time spent executing the program, but I don't have any feel for what real BPF programs look like or have any benchmark suite to run. Is there something out there this should be benchmarked against? (I'd also like to know that to run those benchmarks on the RISC-V port.)
If you run the selftests 'test_progs' with -v it'll measure/print the execution time of the programs. I'd say *most* BPF program invokes a helper (via call). It would be interesting to see, for say the selftests, how often the optimization can be performed.
- Is this the sort of thing that makes sense in a BPF JIT? I guess I've just realized I turned "review this patch" into a way bigger rabbit hole than I really want to go down...
I'd say 'yes'. My hunch, and the workloads I've seen, BPF programs are usually loaded, and then resident for a long time. So, the JIT time is not super critical. The FB/Cilium folks can definitely provide a better sample point, than my hunch. ;-)
In our case the JIT time can be relevant because we are effectively holding up a kubernetes pod load waiting for programs to load. However, we can probably work-around it by doing more aggressive dynamic linking now that this is starting to land.
It would be interesting to have a test to measure load time in selftests or selftests/benchmark/ perhaps. We have some of these out of tree we could push in I think if there is interest.
Björn