On Tue, Jun 18, 2013 at 10:08 PM, Benjamin LaHaise <bcrl@xxxxxxxxx> wrote:
> On Tue, Jun 18, 2013 at 10:05:22PM +0300, Octavian Purdila wrote:
>> On Fri, Jun 14, 2013 at 5:20 PM, Octavian Purdila
>> <octavian.purdila@xxxxxxxxx> wrote:
>> > I picked up Ben's patch and incorporated Zach's idea and the first
>> > results look promising, as expected. I am going to do a full test with
>> > the same workload I've used for rbtree and come back with the results
>> > and the patch in a day or so.
>> >
>>
>> Unfortunately, I still see performance degradation for the one ioctx
>> case even when using IDR. I am using the same fio benchmark as before.
>
> How much of a regression?
>
Hi Ben,
It is even worse then radix, I've seen between 1% to even 4%
degradation. I have tried to profile it and so far the culprit seems
to be the hint code in idr_find. perf top -e dTLB-load-misses shows:
â static inline void *idr_find(struct idr *idr, int id)
â {
â struct idr_layer *hint = rcu_dereference_raw(idr->hint);
â
â if (hint && (id & ~IDR_MASK) == hint->prefix)
â return rcu_dereference_raw(hint->ary[id &
IDR_MASK]);
49.42 â d0: movzbl %r12b,%edx
â add $0x4,%rdx
â mov 0x8(%rax,%rdx,8),%rax
I should say that I am using kvm for testing this. Hopefully this is
not a side effect of virtualization. I have attached the patch I am
testing with.
Thanks,
Tavi
Attachment:
0001-aio-use-ring-id-and-IDR-to-speed-up-lookps.patch
Description: Binary data
