On Wed, 19 May 2021 19:57:55 +0200 Nicolas Saenz Julienne <nsaenzju@xxxxxxxxxx> wrote: > To minimize trace's effect on isolated CPUs. That is, CPUs were only a > handful or a single, process are allowed to run. Introduce a new trace > option: 'poll-rb'. > > This option changes the heuristic used to wait for data on trace > buffers. The default one, based on wait queues, will trigger an IPI[1] > on the CPU responsible for new data, which will take care of waking up > the trace gathering process (generally trace-cmd). Whereas with > 'poll-rb' we will poll (as in busy-wait) the ring buffers from the trace > gathering process, releasing the CPUs writing trace data from doing any > wakeup work. > > This wakeup work, although negligible in the vast majority of workloads, > may cause unwarranted latencies on systems running trace on isolated > CPUs. This is made worse on PREEMPT_RT kernels, as they defer the IPI > handling into a kernel thread, forcing unwarranted context switches on > otherwise extremely busy CPUs. > > To illustrate this, tracing with PREEMPT_RT=y on an isolated CPU with a > single process pinned to it (NO_HZ_FULL=y, and plenty more isolation > options enabled). I see: > - 50-100us latency spikes with the default trace-cmd options > - 14-10us latency spikes with 'poll-rb' > - 11-8us latency spikes with no tracing at all > > The obvious drawback of 'poll-rb' is putting more pressure on the > housekeeping CPUs. Wasting cycles. Hence the notice in the documentation > discouraging its use in general. > > [1] The IPI, in this case, an irq_work, is needed since trace might run > in NMI context. Which is not suitable for wake-ups. Can't this simply be done in user-space? Set the reading of the trace buffers to O_NONBLOCK and it wont wait for buffering to happen, and should prevent it from causing the IPI wake ups. If you need this for trace-cmd, we can add a --poll option that would do this. -- Steve
