Fitting time slices to execution phases

By: Linus Torvalds (torvalds.delete@this.linux-foundation.org), August 21, 2018 3:31 pm
Room: Moderated Discussions
Linus Torvalds (torvalds.delete@this.linux-foundation.org) on August 21, 2018 2:34 pm wrote:
>
> I Again, rseq is designed for per-cpu lockless algorithms. The main use case is
> literally a percpu memory allocator that doesn't need to take locks. There are others, but
> memory allocators can be so critical that you might as well see that as the primary one.

Actually, another thing worth mentioning is just "statistics".

It sounds ridiculously easy to do things like "count number of packets processed" or "count hash misses vs hits", but it turns out that in heavily threaded environments, doing things like operation statistics is sometimes some of the nastiest stuff there is.

It's trivial to do per-thread: just increment the counter in the per-thread area. Boom. Done. There are no races, because even on the worst kind of load-store machine you can just do

get thread pointer
load counter from memory off thread pointer
increment counter
store new counter off thread pointer

and you're done. No locks, no cache conflicts with other threads, no nothing.

The problem with the per-thread model is that it can be very expensive indeed to add all those per-thread statistics up. Thousands of threads is not even unusual. So maybe you can deal with some slop, and just have the threads periodically empoty their per-thread statistics into some cache-coherent pool or something.

It's solvable, don't get me wrong. But it has problems.

So you can solve it other ways. One way is to have the remote atomics, and just do the statistics update "somewhere else" and asynchronously. That's a lovely model in many ways, but it's not portable.

Or you can just do actual cache-coherent locked loads to a global counter. That works fine for most cases, but it gets to be a complete disaster for the really hot counters on big machines. You'll potentially spend more time doing the statistics than you spend on the work you did.

Or you can play games. You could reserve one register for a "percpu base" - the same way you already have a base pointer for the thread local area, and rely on the OS updating it, and if you have a locally atomic (ie not SMP) instruction to do a read-modify-write, you can just do a single instruction
increment statistics(%cpubase)

and you're fine. On x86 you could use a segment register for the percpu base, the same way you use a segment register for the thread-local storage, and you're done.

With the above you get great cache behavior, no cross-cpu migration at all, and much better scaling too - you can add up the statistics by just iterating over the number of CPU's. Sure, that might still be a reasonably big number on a studly machine, but it's a much better proposition than the per-thread case.

But the above already requires that OS support for per-cpu base updates as you move around, and it also requires that you do have that single-instruction read-modify-write. So in practice, you can't actually do it, because it is so unportable.

So this is the kind of thing rseq can help with. A single memory add to update statistics - it should be really trivial and easy, but it actually is anything but. It's complex enough that people don't even do it at all.

With rseq, you can do basically the above "portably" (you have to have rseq) with two normal reads and write instruciton. Regular reads and writes too, not some slow cache-atomic ones.

There's a trivial (again normal instrucvtion) sequence to start the critical section (to a magic location you had registered with the kernel earlier), one read of the percpu base value, one read of the actual counter value, and one write of the new counter value.

If you got scheduled away, you get aborted and restart the sequence (otr maybe you decide to fall back to cacheline atomic, that's probably the better model anyway)

That kind of silly low-level thing that sounds trivial - but really isn't - is what rseq is about. Things that sound so simple that you'd not even realize that they are hard to implement.

So per-cpu memory allocators, trace buffers, counters. Things that sound obvious and trivial, but are pretty basic and aren't trivial at all in user space, and where locking is much too expensive, and transactional memory doesn't even help.

Linus
< Previous Post in ThreadNext Post in Thread >
TopicPosted ByDate
ARM turns to a god and a heroAM2018/08/16 09:32 AM
  ARM turns to a god and a heroMaynard Handley2018/08/16 09:41 AM
    ARM turns to a god and a heroDoug S2018/08/16 11:11 AM
    ARM turns to a god and a heroGeoff Langdale2018/08/16 11:59 PM
      ARM turns to a god and a herodmcq2018/08/17 05:12 AM
  ARM is somewhat misleadingAdrian2018/08/16 11:56 PM
    It's marketing materialGabriele Svelto2018/08/17 01:00 AM
      It's marketing materialMichael S2018/08/17 03:13 AM
        It's marketing materialdmcq2018/08/17 05:23 AM
          It's marketing materialAndrei Frumusanu2018/08/17 07:25 AM
        It's marketing materialLinus Torvalds2018/08/17 11:20 AM
          It's marketing materialGroo2018/08/17 01:44 PM
            It's marketing materialDoug S2018/08/17 02:14 PM
          promises and deliveriesAM2018/08/17 02:32 PM
            promises and deliveriesPassing Through2018/08/17 03:02 PM
              Just by way of clarification Passing Through2018/08/17 03:15 PM
                Just by way of clarification AM2018/08/18 12:49 PM
                  Just by way of clarification Passing Through2018/08/18 01:34 PM
                    This ain't the nineties any longerPassing Through2018/08/18 01:54 PM
                      This ain't the nineties any longerMaynard Handley2018/08/18 02:50 PM
                        This ain't the nineties any longerPassing Through2018/08/18 03:57 PM
                          This ain't the nineties any longerPassing Through2018/09/06 02:42 PM
                            This ain't the nineties any longerMaynard Handley2018/09/07 04:10 PM
                              This ain't the nineties any longerPassing Through2018/09/07 04:48 PM
                                This ain't the nineties any longerMaynard Handley2018/09/07 05:22 PM
                Just by way of clarification Wilco2018/08/18 01:26 PM
                  Just by way of clarification Passing Through2018/08/18 01:39 PM
                  Just by way of clarification none2018/08/18 10:52 PM
                    Just by way of clarification dmcq2018/08/19 08:32 AM
                      Just by way of clarification none2018/08/19 08:54 AM
                        Just by way of clarification dmcq2018/08/19 11:24 AM
                          Just by way of clarification none2018/08/19 11:52 AM
                  Just by way of clarification Gabriele Svelto2018/08/19 06:41 AM
                    Just by way of clarification Passing Through2018/08/19 09:25 AM
                      Whiteboards at Gatwick airport anyone? Passing Through2018/08/20 04:24 AM
          It's marketing materialMichael S2018/08/18 11:12 AM
          It's marketing materialBrett2018/08/18 05:22 PM
            It's marketing materialBrett2018/08/18 05:33 PM
              It's marketing materialAdrian2018/08/19 01:21 AM
        A76AM2018/08/17 02:45 PM
          A76Michael S2018/08/18 11:20 AM
            A76AM2018/08/18 12:39 PM
              A76Michael S2018/08/18 12:49 PM
                A76AM2018/08/18 01:06 PM
                  A76Doug S2018/08/18 01:43 PM
                    A76Maynard Handley2018/08/18 02:42 PM
                      A76Maynard Handley2018/08/18 04:22 PM
                        Why write zeros when one can use metadata?Paul A. Clayton2018/08/18 06:19 PM
                          Why write zeros when one can use metadata?Maynard Handley2018/08/19 11:12 AM
                            Dictionary compress might apply to memcopyPaul A. Clayton2018/08/19 01:45 PM
                        Instructions for zeroingKonrad Schwarz2018/08/30 06:37 AM
                          Instructions for zeroingMaynard Handley2018/08/30 08:41 AM
                          Instructions for zeroingAdrian2018/08/30 11:37 AM
                            dcbz -> dcbzl (was: Instructions for zeroing)hobold2018/08/31 01:50 AM
                              dcbz -> dcbzl (was: Instructions for zeroing)dmcq2018/09/01 05:28 AM
                      A76Travis2018/08/19 11:36 AM
                        A76Maynard Handley2018/08/19 12:22 PM
                          A76Travis2018/08/19 02:07 PM
                            A76Maynard Handley2018/08/19 06:24 PM
                        Remote atomicsmatthew2018/08/19 12:51 PM
                          Remote atomicsMichael S2018/08/19 01:58 PM
                            Remote atomicsmatthew2018/08/19 02:32 PM
                              Remote atomicsMichael S2018/08/19 02:36 PM
                                Remote atomicsmatthew2018/08/19 02:48 PM
                                  Remote atomicsMichael S2018/08/19 03:16 PM
                                    Remote atomicsRicardo B2018/08/20 10:05 AM
                            Remote atomicsdmcq2018/08/19 02:33 PM
                          Remote atomicsTravis2018/08/19 02:32 PM
                            Remote atomicsMichael S2018/08/19 02:46 PM
                              Remote atomicsTravis2018/08/19 05:35 PM
                                Remote atomicsMichael S2018/08/20 03:29 AM
                            Remote atomicsmatthew2018/08/19 07:58 PM
                              Remote atomicsanon2018/08/20 12:59 AM
                                Remote atomicsTravis2018/08/20 10:26 AM
                              Remote atomicsTravis2018/08/20 09:57 AM
                              Remote atomicsLinus Torvalds2018/08/20 04:29 PM
                                Fitting time slices to execution phasesPaul A. Clayton2018/08/21 09:09 AM
                                  Fitting time slices to execution phasesLinus Torvalds2018/08/21 02:34 PM
                                    Fitting time slices to execution phasesLinus Torvalds2018/08/21 03:31 PM
                                      Fitting time slices to execution phasesGabriele Svelto2018/08/21 03:54 PM
                                        Fitting time slices to execution phasesLinus Torvalds2018/08/21 04:26 PM
                                      Fitting time slices to execution phasesTravis2018/08/21 04:21 PM
                                        Fitting time slices to execution phasesLinus Torvalds2018/08/21 04:39 PM
                                          Fitting time slices to execution phasesTravis2018/08/21 04:59 PM
                                            Fitting time slices to execution phasesLinus Torvalds2018/08/21 05:13 PM
                                      Fitting time slices to execution phasesanon2018/08/21 04:27 PM
                                        Fitting time slices to execution phasesLinus Torvalds2018/08/21 06:02 PM
                                          Fitting time slices to execution phasesEtienne2018/08/22 02:28 AM
                                        Fitting time slices to execution phasesGabriele Svelto2018/08/22 03:07 PM
                                          Fitting time slices to execution phasesTravis2018/08/22 04:00 PM
                                          Fitting time slices to execution phasesanon2018/08/22 06:52 PM
                                    Fitting time slices to execution phasesTravis2018/08/21 04:37 PM
                                    Is preventing misuse that complex?Paul A. Clayton2018/08/23 05:42 AM
                                      Is preventing misuse that complex?Linus Torvalds2018/08/23 12:46 PM
                                        Is preventing misuse that complex?Travis2018/08/23 01:29 PM
                                          Is preventing misuse that complex?Travis2018/08/23 01:33 PM
                                            Is preventing misuse that complex?Jeff S.2018/08/24 07:57 AM
                                              Is preventing misuse that complex?Travis2018/08/24 08:47 AM
                                          Is preventing misuse that complex?Linus Torvalds2018/08/23 02:30 PM
                                            Is preventing misuse that complex?Travis2018/08/23 03:11 PM
                                              Is preventing misuse that complex?Linus Torvalds2018/08/24 01:00 PM
                                                Is preventing misuse that complex?Gabriele Svelto2018/08/24 01:25 PM
                                                  Is preventing misuse that complex?Linus Torvalds2018/08/24 01:33 PM
                                  Fitting time slices to execution phasesTravis2018/08/21 03:54 PM
                                rseq: holy grail rwlock?Travis2018/08/21 03:18 PM
                                  rseq: holy grail rwlock?Linus Torvalds2018/08/21 03:59 PM
                                    rseq: holy grail rwlock?Travis2018/08/21 04:27 PM
                                      rseq: holy grail rwlock?Linus Torvalds2018/08/21 05:10 PM
                                        rseq: holy grail rwlock?Travis2018/08/21 06:21 PM
                  ARM design housesMichael S2018/08/21 05:07 AM
                    ARM design housesWilco2018/08/22 12:38 PM
                      ARM design housesMichael S2018/08/22 02:21 PM
                        ARM design housesWilco2018/08/22 03:23 PM
                          ARM design housesMichael S2018/08/29 01:58 AM
                            Qualcomm's core naming scheme really, really sucksHeikki Kultala2018/08/29 02:19 AM
                A76Maynard Handley2018/08/18 02:07 PM
                  A76Michael S2018/08/18 02:32 PM
                    A76Maynard Handley2018/08/18 02:52 PM
                      A76Michael S2018/08/18 03:04 PM
    ARM is somewhat misleadingjuanrga2018/08/17 01:20 AM
    Surprised??Alberto2018/08/17 01:52 AM
      Surprised??Alberto2018/08/17 02:10 AM
      Surprised??none2018/08/17 02:46 AM
      Garbage talkAndrei Frumusanu2018/08/17 07:30 AM
        Garbage talkMichael S2018/08/17 07:43 AM
          Garbage talkAndrei Frumusanu2018/08/17 09:51 AM
            Garbage talkMichael S2018/08/18 11:29 AM
        Garbage talkAdrian2018/08/17 08:28 AM
          Garbage talkAlberto2018/08/17 09:20 AM
          Garbage talkAndrei Frumusanu2018/08/17 09:48 AM
            Garbage talkAdrian2018/08/17 10:17 AM
              Garbage talkAndrei Frumusanu2018/08/17 10:36 AM
                Garbage talkAdrian2018/08/17 02:53 PM
                  Garbage talkAndrei Frumusanu2018/08/18 12:17 AM
        More like a religion he?? ARM has an easy life :)Alberto2018/08/17 09:13 AM
          More like a religion he?? ARM has an easy life :)Andrei Frumusanu2018/08/17 09:34 AM
            More like a religion he?? ARM has an easy life :)Alberto2018/08/17 10:03 AM
              More like a religion he?? ARM has an easy life :)Andrei Frumusanu2018/08/17 10:43 AM
              More like a religion he?? ARM has an easy life :)Doug S2018/08/17 02:17 PM
              15W phone SoCsAM2018/08/17 03:04 PM
          More like a religion he?? ARM has an easy life :)Maynard Handley2018/08/17 12:29 PM
  my future stuff will be better than your old stuff, hey I'm a god at last (NT)Eric Bron2018/08/18 03:34 AM
    my future stuff will be better than your old stuff, hey I'm a god at lastnone2018/08/18 08:34 AM
Reply to this Topic
Name:
Email:
Topic:
Body: No Text
How do you spell avocado?