On Fri, Nov 22, 2019 at 06:33:25PM -0800, Matthew Wilcox wrote: > On Sat, Nov 23, 2019 at 10:16:39AM +0800, zhengbin (A) wrote: > > By the way, percpu IDA is for reducing performance impact? This patch has 2.16% > > performance degradation(Use perf to get the cost of ida_alloc_range) > > 2.16% degradation in your tests ... I bet Eric didn't make it so complex > for only 2% performance impact. Unfortunately, he didn't give any > numbers in his patch submission, but it's going to be a bigger problem > on multi-socket machines than on a laptop. It could also be that socket(2) is called (on real loads) just a bit more often than open(2) in a benchmark that tests file creation...
