Message-ID: <BA750BC0-0A3B-488B-806C-90C1B6CDF586@fb.com>
Date:   Wed, 5 Aug 2020 07:01:00 +0000
From:   Song Liu <songliubraving@...com>
To:     Andrii Nakryiko <andrii.nakryiko@...il.com>
CC:     open list <linux-kernel@...r.kernel.org>,
        bpf <bpf@...r.kernel.org>, Networking <netdev@...r.kernel.org>,
        Alexei Starovoitov <ast@...nel.org>,
        Daniel Borkmann <daniel@...earbox.net>,
        Kernel Team <Kernel-team@...com>,
        john fastabend <john.fastabend@...il.com>,
        KP Singh <kpsingh@...omium.org>,
        Jesper Dangaard Brouer <brouer@...hat.com>,
        Daniel Xu <dlxu@...com>
Subject: Re: [PATCH bpf-next 5/5] selftests/bpf: add benchmark for uprobe vs.
 user_prog



> On Aug 4, 2020, at 10:47 PM, Andrii Nakryiko <andrii.nakryiko@...il.com> wrote:
> 
> On Tue, Aug 4, 2020 at 9:47 PM Song Liu <songliubraving@...com> wrote:
>> 
>> 
>> 
>>> On Aug 4, 2020, at 6:52 PM, Andrii Nakryiko <andrii.nakryiko@...il.com> wrote:
>>> 
>>> On Tue, Aug 4, 2020 at 2:01 PM Song Liu <songliubraving@...com> wrote:
>>>> 
>>>> 
>>>> 
>>>>> On Aug 2, 2020, at 10:10 PM, Andrii Nakryiko <andrii.nakryiko@...il.com> wrote:
>>>>> 
>>>>> On Sun, Aug 2, 2020 at 9:47 PM Song Liu <songliubraving@...com> wrote:
>>>>>> 
>>>>>> 
>>>>>>> On Aug 2, 2020, at 6:51 PM, Andrii Nakryiko <andrii.nakryiko@...il.com> wrote:
>>>>>>> 
>>>>>>> On Sat, Aug 1, 2020 at 1:50 AM Song Liu <songliubraving@...com> wrote:
>>>>>>>> 
>>>>>>>> Add a benchmark to compare performance of
>>>>>>>> 1) uprobe;
>>>>>>>> 2) user program w/o args;
>>>>>>>> 3) user program w/ args;
>>>>>>>> 4) user program w/ args on random cpu.
>>>>>>>> 
>>>>>>> 
>>>>>>> Can you please add it to the existing benchmark runner instead, e.g.,
>>>>>>> along the other bench_trigger benchmarks? No need to re-implement
>>>>>>> benchmark setup. And also that would also allow to compare existing
>>>>>>> ways of cheaply triggering a program vs this new _USER program?
>>>>>> 
>>>>>> Will try.
>>>>>> 
>>>>>>> 
>>>>>>> If the performance is not significantly better than other ways, do you
>>>>>>> think it still makes sense to add a new BPF program type? I think
>>>>>>> triggering KPROBE/TRACEPOINT from bpf_prog_test_run() would be very
>>>>>>> nice, maybe it's possible to add that instead of a new program type?
>>>>>>> Either way, let's see comparison with other program triggering
>>>>>>> mechanisms first.
>>>>>> 
>>>>>> Triggering KPROBE and TRACEPOINT from bpf_prog_test_run() will be useful.
>>>>>> But I don't think they can be used instead of user program, for a couple
>>>>>> reasons. First, KPROBE/TRACEPOINT may be triggered by other programs
>>>>>> running in the system, so user will have to filter those noise out in
>>>>>> each program. Second, it is not easy to specify CPU for KPROBE/TRACEPOINT,
>>>>>> while this feature could be useful in many cases, e.g. get stack trace
>>>>>> on a given CPU.
>>>>>> 
>>>>> 
>>>>> Right, it's not as convenient with KPROBE/TRACEPOINT as with the USER
>>>>> program you've added specifically with that feature in mind. But if
>>>>> you pin user-space thread on the needed CPU and trigger kprobe/tp,
>>>>> then you'll get what you want. As for the "noise", see how
>>>>> bench_trigger() deals with that: it records thread ID and filters
>>>>> everything not matching. You can do the same with CPU ID. It's not as
>>>>> automatic as with a special BPF program type, but still pretty simple,
>>>>> which is why I'm still deciding (for myself) whether USER program type
>>>>> is necessary :)
>>>> 
>>>> Here are some bench_trigger numbers:
>>>> 
>>>> base      :    1.698 ± 0.001M/s
>>>> tp        :    1.477 ± 0.001M/s
>>>> rawtp     :    1.567 ± 0.001M/s
>>>> kprobe    :    1.431 ± 0.000M/s
>>>> fentry    :    1.691 ± 0.000M/s
>>>> fmodret   :    1.654 ± 0.000M/s
>>>> user      :    1.253 ± 0.000M/s
>>>> fentry-on-cpu:    0.022 ± 0.011M/s
>>>> user-on-cpu:    0.315 ± 0.001M/s
>>>> 
>>> 
>>> Ok, so basically all of raw_tp,tp,kprobe,fentry/fexit are
>>> significantly faster than USER programs. Sure, when compared to
>>> uprobe, they are faster, but not when doing on-specific-CPU run, it
>>> seems (judging from this patch's description, if I'm reading it
>>> right). Anyways, speed argument shouldn't be a reason for doing this,
>>> IMO.
>>> 
>>>> The two "on-cpu" tests run the program on a different CPU (see the patch
>>>> at the end).
>>>> 
>>>> "user" is about 25% slower than "fentry". I think this is mostly because
>>>> getpgid() is a faster syscall than bpf(BPF_TEST_RUN).
>>> 
>>> Yes, probably.
>>> 
>>>> 
>>>> "user-on-cpu" is more than 10x faster than "fentry-on-cpu", because IPI
>>>> is way faster than moving the process (via sched_setaffinity).
>>> 
>>> I don't think that's a good comparison, because you are actually
>>> testing sched_setaffinity performance on each iteration vs IPI in the
>>> kernel, not a BPF overhead.
>>> 
>>> I think the fair comparison for this would be to create a thread and
>>> pin it on necessary CPU, and only then BPF program calls in a loop.
>>> But I bet any of existing program types would beat USER program.
>>> 
>>>> 
>>>> For use cases that we would like to call BPF program on specific CPU,
>>>> triggering it via IPI is a lot faster.
>>> 
>>> So these use cases would be nice to expand on in the motivational part
>>> of the patch set. It's not really emphasized and it's not at all clear
>>> what you are trying to achieve. It also seems, depending on latency
>>> requirements, it's totally possible to achieve comparable results by
>>> pre-creating a thread for each CPU, pinning each one to its designated
>>> CPU and then using any suitable user-space signaling mechanism (a
>>> queue, condvar, etc) to ask a thread to trigger BPF program (fentry on
>>> getpgid(), for instance).
>> 
>> I don't see why user space signal plus fentry would be faster than IPI.
>> If the target cpu is running something, this gonna add two context
>> switches.
>> 
> 
> I didn't say faster, did I? I said it would be comparable and wouldn't
> require a new program type.

Well, I don't think adding program type is that big a deal. If that is
really a problem, we can use a new attach type instead. The goal is to 
trigger it with sys_bpf() on a different cpu. So we can call it kprobe
attach to nothing and hack that way. I add the new type because it makes
sense. The user just want to trigger a BPF program from user space. 

> But then again, without knowing all the
> details, it's a bit hard to discuss this. E.g., if you need to trigger
> that BPF program periodically, you can sleep in those per-CPU threads,
> or epoll, or whatever. Or maybe you can set up a per-CPU perf event
> that would trigger your program on the desired CPU, etc.My point is
> that I and others shouldn't be guessing this, I'd expect someone who's
> proposing an entire new BPF program type to motivate why this new
> program type is necessary and what problem it's solving that can't be
> solved with existing means.

Yes, there are other options. But they all come with non-trivial cost. 
Per-CPU-per-process threads and/or per-CPU perf event are cost we have 
to pay in production. IMO, these cost are much higher than a new program
type (or attach type). 

> 
> BTW, how frequently do you need to trigger the BPF program? Seems very
> frequently, if 2 vs 1 context switches might be a problem?

The whole solution requires two BPF programs. One on each context switch, 
the other is the user program. The user program will not trigger very
often. 

> 
>>> I bet in this case the  performance would be
>>> really nice for a lot of practical use cases. But then again, I don't
>>> know details of the intended use case, so please provide some more
>>> details.
>> 
>> Being able to trigger BPF program on a different CPU could enable many
>> use cases and optimizations. The use case I am looking at is to access
>> perf_event and percpu maps on the target CPU. For example:
>>        0. trigger the program
>>        1. read perf_event on cpu x;
>>        2. (optional) check which process is running on cpu x;
>>        3. add perf_event value to percpu map(s) on cpu x.
>> 
>> If we do these steps in a BPF program on cpu x, the cost is:
>>        A.0) trigger BPF via IPI;
>>        A.1) read perf_event locally;
>>        A.2) local access current;
>>        A.3) local access of percpu map(s).
>> 
>> If we can only do these on a different CPU, the cost will be:
>>        B.0) trigger BPF locally;
>>        B.1) read perf_event via IPI;
>>        B.2) remote access current on cpu x;
>>        B.3) remote access percpu map(s), or use non-percpu map(2).
>> 
>> Cost of (A.0 + A.1) is about same as (B.0 + B.1), maybe a little higher
>> (sys_bpf(), vs. sys_getpgid()). But A.2 and A.3 will be significantly
>> cheaper than B.2 and B.3.
>> 
>> Does this make sense?
> 
> It does, thanks. But what I was describing is still A, no? BPF program
> will be triggered on your desired cpu X, wouldn't it?

Well, that would be option C, but C could not do step 2, because we context 
switch to the dedicated thread. 

Powered by blists - more mailing lists