[Pmem] Re: [PATCH] anolis: mm: Add page promotion throttle statistic

Baolin Wang <baolin.wang(a)linux.alibaba.com> writes: > ( 2/23/2022 4:44 PM, Huang, Ying S: >> Baolin Wang <baolin.wang(a)linux.alibaba.com> writes: >> >>> ( 2/23/2022 3:51 PM, Huang, Ying S: >>>> Baolin Wang <baolin.wang(a)linux.alibaba.com> writes: >>>> >>>>> ANBZ: #80 >>>>> >>>>> Add page promotion throttle statistic, which can be used to check >>>>> how many cold pages were trying to be prmoted to DRAM, and help >>>>> to tuning the latency threhold. >>>> Can this be calculated via the following formula? >>>> numa_hint_faults - pgpromote_candidate >>> >>> They are not same. Since numa_hint_faults will not contain file page >>> promotion statistics, while the should_numa_migrate_memory() will >>> check the file pages latency before promotion. Here some statistics I >>> observed: >>> >>> pgpromote_candidate 3890100 >>> pgpromote_cold_throttle 50342001 >>> numa_hint_faults 35341839 >> Then how about adding a counter for unmapped pages accesses we >> checked? >> And I think these kind of counters are for debugging only at least >> for >> now. Should we merge them formally? >> Do you have an example about how to use this new counter or similar >> one? > > We can add a counter for unmapped pages, but it still be little clear > for these cold throttling counter. Please also considering > numa_hint_faults will contain local faults, and some page promotion > counts if the DRAM is free enough. > > I think the cold throttling counter is not only for debugging, but > also for doing some decision in future. We can know how many cold > pages are accessed, and we can measure the cold and hot memory > distribution for this workdload to decide if the workload is suitable > for the tiered-memory. So a clear and accurate cold throttling counter > will be helpful. How do you think? If our target is the page temperature distribution, then a histogram may be even better. I have implemented one before. The histogram will count the hint page fault number within each access latency range. And we can output and clear the histogram every sysctl_numa_balancing_scan_period_max.