sem_otime trashing
Hi Rik,
I finally managed to get EFI boot, i.e. I'm now able to test on my i3 (2core+HT).
With semscale (i.e.: just overhead, perform semop=0 operations), the scalability from 1 to 2 cores is good, but not linear:
# semscale 10 | grep "interleave 2"
Quoted 2 lines
Cpus 1, interleave 2 delay 0: 35502103 in 10 secs
Cpus 2, interleave 2 delay 0: 53990954 in 10 secs
---
+53% when adding the 2nd core
(interleave 2 to force to use different cores)
Did you consider moving sem_otime into the individual semaphores? I did that (gross patch attached), and the performance is significantly better:
# semscale 10 | grep "interleave 2"
Cpus 1, interleave 2 delay 0: 35585634 in 10 secs
Cpus 2, interleave 2 delay 0: 70410230 in 10 secs
---
+99% scalability when adding the 2nd coreUnfortunately I won't be able to read my mails next week, but the effect was too significant not to share it immediately.
--
Manfred
Makefile +1 -1
@@ -1,7 +1,7 @@ VERSION = 3 PATCHLEVEL = 10 SUBLEVEL = 0 -EXTRAVERSION = -rc3 +EXTRAVERSION = -rc3-otime NAME = Unicycling Gorilla # *DOCUMENTATION*
include/linux/sem.h +0 -1
@@ -12,7 +12,6 @@ struct task_struct;struct sem_array {struct kern_ipc_perm ____cacheline_aligned_in_smp sem_perm;/* permissions .. see ipc.h */ -time_t sem_otime;/* last semop time */ time_t sem_ctime;/* last change time */ struct sem *sem_base;/* ptr to first semaphore in array */ struct list_head pending_alter;/* pending operations */
ipc/sem.c +33 -7
@@ -92,6 +92,7 @@/* One semaphore structure for each semaphore in the system. */ struct sem { +char filler[64 ];int semval;/* current value */ int sempid;/* pid of last operation */ spinlock_t lock;/* spinlock for fine-grained semtimedop */
@@ -99,7 +100,8 @@ struct sem {/* that alter the semaphore */ struct list_head pending_const;/* pending single-sop operations */ /* that do not alter the semaphore*/ -}; +time_t sem_otime;/* candidate for sem_otime */ +}____cacheline_aligned_in_smp ;/* One queue for each sleeping process in the system. */ struct sem_queue {
@@ -919,8 +921,14 @@ static void do_smart_update(struct sem_array *sma, struct sembuf *sops, int nsop } } } -if (otime) - sma->sem_otime =get_seconds (); +if (otime) { +if (sops ==NULL ) { + sma->sem_base[0 ].sem_otime =get_seconds (); + }else { + sma->sem_base[sops[0 ].sem_num].sem_otime = +get_seconds (); + } + } }
@@ -1066,6 +1074,21 @@ static unsigned long copy_semid_to_user(void __user *buf, struct semid64_ds *in, } } +static time_t get_semotime (struct sem_array *sma) +{ +int i; +time_t res; + + res = sma->sem_base[0 ].sem_otime; +for (i =1 ; i < sma->sem_nsems; i++) { +time_t to = sma->sem_base[i].sem_otime; + +if (to > res) + res = to; + } +return res; +} +static int semctl_nolock (struct ipc_namespace *ns,int semid,int cmd,int version,void __user *p) {
@@ -1139,9 +1162,9 @@ static int semctl_nolock(struct ipc_namespace *ns, int semid, goto out_unlock;kernel_to_ipc64_perm (&sma->sem_perm, &tbuf.sem_perm); - tbuf.sem_otime = sma->sem_otime; - tbuf.sem_ctime = sma->sem_ctime; - tbuf.sem_nsems = sma->sem_nsems; + tbuf.sem_otime =get_semotime (sma); + tbuf.sem_ctime = sma->sem_ctime; + tbuf.sem_nsems = sma->sem_nsems;rcu_read_unlock ();if (copy_semid_to_user (p, &tbuf, version))return -EFAULT ;
@@ -2029,6 +2052,9 @@ static int sysvipc_sem_proc_show(struct seq_file *s, void *it) {struct user_namespace *user_ns =seq_user_ns (s);struct sem_array *sma = it; +time_t sem_otime; + + sem_otime =get_semotime (sma); return seq_printf(s, "%10d %10d %4o %10u %5u %5u %5u %5u %10lu %10lu\n",
@@ -2040,7 +2066,7 @@ static int sysvipc_sem_proc_show(struct seq_file *s, void *it)from_kgid_munged (user_ns, sma->sem_perm.gid),from_kuid_munged (user_ns, sma->sem_perm.cuid),from_kgid_munged (user_ns, sma->sem_perm.cgid), - sma->sem_otime, + sem_otime, sma->sem_ctime); }#endif
Quoted 18 lines · Manfred Spraul
Hi Rik,
I finally managed to get EFI boot, i.e. I'm now able to test on my i3 (2core+HT).
With semscale (i.e.: just overhead, perform semop=0 operations), the scalability from 1 to 2 cores is good, but not linear:
# semscale 10 | grep "interleave 2"Quoted 2 lines
Cpus 1, interleave 2 delay 0: 35502103 in 10 secs
Cpus 2, interleave 2 delay 0: 53990954 in 10 secs---
+53% when adding the 2nd core
(interleave 2 to force to use different cores)Did you consider moving
sem_otimeinto the individual semaphores? I did that (gross patch attached), and the performance is significantly better:
Quoted 8 lines
# semscale 10 | grep "interleave 2" Cpus 1, interleave 2 delay 0: 35585634 in 10 secs Cpus 2, interleave 2 delay 0: 70410230 in 10 secs --- +99% scalability when adding the 2nd coreUnfortunately I won't be able to read my mails next week, but the effect was too significant not to share it immediately.
64 core box.
Previous numbers:
vogelweide:/abuild/mike/:[0]# uname -r 3.8.13-rt9-rtm vogelweide:/abuild/mike/:[0]# ./semop-multi 256 64 cpus 64, threads: 256, semaphores: 64, test duration: 30 secs total operations: 33553800, ops/sec 1118460
New numbers:
vogelweide:/abuild/mike/:[0]# !./semop-multi ./semop-multi 256 64 cpus 64, threads: 256, semaphores: 64, test duration: 30 secs total operations: 129474934, ops/sec 4315831
But, box rcu stalled on me. It's looking like the scalability patches are a bit racy rcu wise in an -rt kernel (oh dear). So, build as plain old PREEMPT again, eliminate -rt funnies.
Previous numbers:
vogelweide:/abuild/mike/:[0]# ./semop-multi 256 64 cpus 64, threads: 256, semaphores: 64, test duration: 30 secs total operations: 22053968, ops/sec 735132
vogelweide:/abuild/mike/:[0]# ./osim 64 256 1000000 0 0 osim <sems> <tasks> <loops> <busy-in> <busy-out> osim: using a semaphore array with 64 semaphores. osim: using 256 tasks. osim: each thread loops 3907 times osim: each thread busyloops 0 loops outside and 0 loops inside. total execution time: 1.858765 seconds for 1000192 loops per loop execution time: 1.858 usec
New numbers:
vogelweide:/abuild/mike/:[0]# !./semop ./semop-multi 256 64 cpus 64, threads: 256, semaphores: 64, test duration: 30 secs total operations: 45521478, ops/sec 1517382 vogelweide:/abuild/mike/:[0]# !./osim ./osim 64 256 1000000 0 0 osim <sems> <tasks> <loops> <busy-in> <busy-out> osim: using a semaphore array with 64 semaphores. osim: using 256 tasks. osim: each thread loops 3907 times osim: each thread busyloops 0 loops outside and 0 loops inside. total execution time: 0.350682 seconds for 1000192 loops per loop execution time: 0.350 usec
(1.8->0.3?.. box, you ain't a race horse, you're a plow horse)
vogelweide:/abuild/mike/:[0]# ./osim 64 256 1000000 0 0 osim <sems> <tasks> <loops> <busy-in> <busy-out> osim: using a semaphore array with 64 semaphores. osim: using 256 tasks. osim: each thread loops 3907 times osim: each thread busyloops 0 loops outside and 0 loops inside. total execution time: 0.276405 seconds for 1000192 loops per loop execution time: 0.276 usec vogelweide:/abuild/mike/:[0]# ./osim 64 256 1000000 0 0 osim <sems> <tasks> <loops> <busy-in> <busy-out> osim: using a semaphore array with 64 semaphores. osim: using 256 tasks. osim: each thread loops 3907 times osim: each thread busyloops 0 loops outside and 0 loops inside. total execution time: 0.370041 seconds for 1000192 loops per loop execution time: 0.369 usec vogelweide:/abuild/mike/:[0]# ./osim 64 256 1000000 0 0 osim <sems> <tasks> <loops> <busy-in> <busy-out> osim: using a semaphore array with 64 semaphores. osim: using 256 tasks. osim: each thread loops 3907 times osim: each thread busyloops 0 loops outside and 0 loops inside. total execution time: 0.502396 seconds for 1000192 loops per loop execution time: 0.502 usec
(runtime)
vogelweide:/abuild/mike/:[0]# ./osim 64 256 10000000 0 0 osim <sems> <tasks> <loops> <busy-in> <busy-out> osim: using a semaphore array with 64 semaphores. osim: using 256 tasks. osim: each thread loops 39063 times osim: each thread busyloops 0 loops outside and 0 loops inside. total execution time: 3.354423 seconds for 10000128 loops per loop execution time: 0.335 usec vogelweide:/abuild/mike/:[0]# ./osim 64 256 100000000 0 0 osim <sems> <tasks> <loops> <busy-in> <busy-out> osim: using a semaphore array with 64 semaphores. osim: using 256 tasks. osim: each thread loops 390625 times osim: each thread busyloops 0 loops outside and 0 loops inside. total execution time: 41.180479 seconds for 100000000 loops per loop execution time: 0.411 usec
Box likes your idea.