Prefetch kernel stacks to speed up context switch
Repost previously discussed patch (on Jul 27, 2005). Ingo did the same thing for all arch with 471 lines of patch. I'm still advocating this little 30 lines patch, of 6 lines introduces prefetch_stack() generic interface.
Andrew, please consider -mm inclusion. Or advise me what I need to do to take this forward. Thanks.
- Ken
------
For architecture like ia64, the switch stack structure is fairly large (currently 528 bytes). For context switch intensive application, we found that significant amount of cache misses occurs in switch_to() function. The following patch adds a hook in the schedule() function to prefetch switch stack structure as soon as 'next' task is determined. This allows maximum overlap in prefetch cache lines for that structure.
arch/ia64/kernel/entry.S | 23 +++++++++++++++++++++++ include/asm-ia64/system.h | 1 + include/linux/sched.h | 5 +++++ kernel/sched.c | 1 + 4 files changed, 30 insertions(+)
arch/ia64/kernel/entry.S renamed from arch/ia64/kernel/entry.S.orig +23 -0
@@ -470,6 +470,29 @@ ENTRY(load_switch_stack) br.cond.sptk.many b7 END(load_switch_stack) +GLOBAL_ENTRY(prefetch_stack) + add r14 = -IA64_SWITCH_STACK_SIZE, sp + add r15 = IA64_TASK_THREAD_KSP_OFFSET, in0 + ;; + ld8 r16 = [r15] // load next's stack pointer + lfetch.fault.excl [r14], 128 + ;; + lfetch.fault.excl [r14], 128 + lfetch.fault [r16], 128 + ;; + lfetch.fault.excl [r14], 128 + lfetch.fault [r16], 128 + ;; + lfetch.fault.excl [r14], 128 + lfetch.fault [r16], 128 + ;; + lfetch.fault.excl [r14], 128 + lfetch.fault [r16], 128 + ;; + lfetch.fault [r16], 128 + br.ret.sptk.many rp +END(prefetch_switch_stack) + GLOBAL_ENTRY(execve) mov r15=__NR_execve // put syscall number in place break __BREAK_SYSCALL
include/asm-ia64/system.h renamed from include/asm-ia64/system.h.orig +1 -0
@@ -274,6 +274,7 @@ extern void ia64_load_extra (struct task*/ #define __ARCH_WANT_UNLOCKED_CTXSW +#define ARCH_HAS_PREFETCH_SWITCH_STACK #define ia64_platform_is (x) (strcmp(x, platform_name) = 0)void cpu_idle_wait (void );
include/linux/sched.h renamed from include/linux/sched.h.orig +5 -0
@@ -592,6 +592,11 @@ extern int groups_search(struct group_in#define GROUP_AT (gi, i) \ ((gi)->blocks[(i)/NGROUPS_PER_BLOCK][(i)%NGROUPS_PER_BLOCK]) +#ifdef ARCH_HAS_PREFETCH_SWITCH_STACK +extern void prefetch_stack (struct task_struct *); +#else +static inline void prefetch_stack (struct task_struct *t) { } +#endif struct audit_context ;/* See audit.c */ struct mempolicy ;
kernel/sched.c renamed from kernel/sched.c.orig +1 -0
@@ -2887,6 +2887,7 @@ switch_tasks:if (next = rq->idle)schedstat_inc (rq, sched_goidle);prefetch (next); +prefetch_stack (next);clear_tsk_need_resched (prev);rcu_qsctr_inc (task_cpu (prev));
Quoted 9 lines · Kenneth W
Repost previously discussed patch (on Jul 27, 2005). Ingo did the same thing for all arch with 471 lines of patch. I'm still advocating this little 30 lines patch, of 6 lines introduces
prefetch_stack()generic interface.Andrew, please consider -mm inclusion. Or advise me what I need to do to take this forward. Thanks.
- Ken
Do you have any benchmarks?
Signature · 2 lines
Coywolf Qi Hunt http://sosdg.org/~coywolf/
Quoted 2 lines · Coywolf Qi Hunt
Quoted 1 line · Kenneth W
Repost previously discussed patch (on Jul 27, 2005).
For reference:
http://www.gelato.unsw.edu.au/archives/linux-ia64/0507/14686.html
Quoted 1 line
Do you have any benchmarks?
Have you read the discussion?
See:
http://www.gelato.unsw.edu.au/archives/linux-ia64/0507/14685.html
Ken didn't post any benchmark results originally but clearly stated the cacheline misses were occurring enough to be measurable. Just knowing which workload and how much much time was spent stalling for cacheline misses should be sufficient.
grant