Messages in this thread | | | From | "John Hawkes" <> | Subject | Re: [PATCH] IA32 SMP spinlock metering | Date | Mon, 11 Oct 1999 15:46:36 -0700 |
| |
----- Original Message ----- From: Ingo Molnar <mingo@chiara.csoma.elte.hu>
> On Tue, 5 Oct 1999, John Hawkes wrote: > > > My test results: ... > > With this workload on this 4xCPU Xeon hardware, spinlock contention in > > 2.2.10 consumed about 8% of theoretically available CPU cycles (340 > > millisecs/sec of waiting out of the 4,000 millisecs/sec theoretically > > available) vs. about 2% in 2.3.11 (95 millisecs of waiting out of 4,000 > > millisecs). > > > > The kernel_flag usage is still significant in 2.3, but its contention is > > greatly reduced. [...] > > have you checked out 2.3.18/19 already? Compared to 2.3.11 it has fully > threaded VFS, buffer-cache and VM, this should further cut down on > kernel_flag usage. I suspect it should be well below 1% already.
For various reasons I haven't been able to rerun my tests with a 2.3.18/19 kernel on a 4xCPU system; but on a 2xCPU system, tests suggest on the order of a 10% (YMMV) decrease in spinlock "wait time" cycles between 2.3.16 and 2.3.18. That number may be wildly inaccurate, but at least it's hinting that we're headed in the right direction.
-- John Hawkes http://oss.sgi.com
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
| |