On 02/23/2011 12:39 PM, Jan Kiszka wrote: > You should try to trace the event flow in qemu, either via strace, via > the built-in tracer (which likely requires a bit more tracepoints), or > via a system-level tracer (ftrace / kernelshark).
The apparent problem is that 25% of cycles is spent in mutex locking and unlocking. But in fact, the real problem is that 90% of the time is spent doing something else than executing code. QEMU exits _a lot_ due to the vm_clock timers. The deadlines are rarely more than a few ms ahead, and at 1 MIPS that leaves room for executing a few thousand instructions for each context switch. The iothread overhead is what makes the situation so bad, because it takes a lot more time to execute those instructions. We do so many (almost) useless passes through cpu_exec_all that even microoptimization helps, for example this: --- a/cpus.c +++ b/cpus.c @@ -767,10 +767,6 @@ static void qemu_wait_io_event_common(CPUState *env) { CPUState *env; - while (all_cpu_threads_idle()) { - qemu_cond_timedwait(tcg_halt_cond, &qemu_global_mutex, 1000); - } - qemu_mutex_unlock(&qemu_global_mutex); /* @@ -1110,7 +1111,15 @@ bool cpu_exec_all(void) } } exit_request = 0; + +#ifdef CONFIG_IOTHREAD + while (all_cpu_threads_idle()) { + qemu_cond_timedwait(tcg_halt_cond, &qemu_global_mutex, 1000); + } + return true; +#else return !all_cpu_threads_idle(); +#endif } void set_numa_modes(void) is enough to cut all_cpu_threads_idle from 9 to 4.5% (not unexpected: the number of calls is halved). But it shouldn't be that high anyway, so I'm not proposing the patch formally. Additionally, the fact that the execution is 99.99% lockstep means you cannot really overlap any part of the I/O and VCPU threads. I found a couple of inaccuracies in my patches that already cut 50% of the time, though. > Did my patches contribute a bit to overhead reduction? They specifically > target the costly vcpu/iothread switches in TCG mode (caused by TCGs > excessive lock-holding times). Yes, they cut 15%. Paolo