On Wed, Apr 01, 2020 at 09:22:24AM +0800, Yafang Shao wrote: > On Tue, Mar 31, 2020 at 11:11 PM Johannes Weiner <hannes@xxxxxxxxxxx> wrote: > > On Fri, Mar 27, 2020 at 09:17:59AM +0800, Yafang Shao wrote: > > > On Thu, Mar 26, 2020 at 10:31 PM Johannes Weiner <hannes@xxxxxxxxxxx> wrote: > With the newly added facility, we can know when these events occur > and how long each event takes. > Then we can use these datas to generate a Latency Heat Map[1] and to > compare whether these latencies match with the application latencies > recoreded in its log - for example the slow query log in mysql. If the > refault latencies match with the slow query log, then these slow > queries are caused by these workingset refault. If the refault > latencies don't match with slow query log, IOW much smaller than the > slow query log, then the slow query log isn't caused by the > workingset refault. Okay, you want to use it much finer-grained to understand individual end-to-end latencies for your services, rather than "the system is melting down and I want to know why." That sounds valid to me. > > > > Can you elaborate a bit how you are using this information? It's not > > > > quite clear to me from the example in patch #2. > > > > > > > > > > From the traced data in patch #2, we can find that the high latencies > > > of user tasks are always type 7 of memstall , which is > > > MEMSTALL_WORKINGSET_THRASHING, and then we should look into the > > > details of wokingset of the user tasks and think about how to improve > > > it - for example, by reducing the workingset. > > > > That's an analyses we run frequently as well: we see high pressure, > > and then correlate it with the events. > > > > High rate of refaults? The workingset is too big. > > > > High rate of compaction work? Somebody is asking for higher order > > pages under load; check THP events next. > > > > etc. > > > > This works fairly reliably. I'm curious what the extra per-event > > latency breakdown would add and where it would be helpful. > > > > I'm not really opposed to your patches it if it is, I just don't see > > the usecase right now. > > > > As I explained above, the rate of these events can't give us a full > view of the memory pressure. High memory pressure may not caused by > high rate of vmstat events, while it can be caused by low rate of > events but with high latencies. Latencies are the application really > concerns, that's why PSI is very useful for us. > > Furthermore, there're some events are not recored in vmstat. e.g. > > typf of memstall vmstat event > > MEMSTALL_KSWAPD pageoutrun, {pgscan, > pgsteal}_kswapd > MEMSTALL_RECLAIM_DIRECT {pgscan,steal}_direct > MEMSTALL_RECLAIM_MEMCG /* no event */ > MEMSTALL_RECLAIM_HIGH /* no event */ > MEMSTALL_KCOMPACTD compact_daemon_wake > MEMSTALL_COMPACT compact_{stall, fail, success} > MEMSTALL_WORKINGSET_REFAULT workingset_refault > MEMSTALL_WORKINGSET_THRASH /* no event */ > MEMSTALL_MEMDELAY /* no event */ > MEMSTALL_SWAPIO pswpin > > After we add these types of memstall, we don't need to add these > missed events one by one. I'm a bit concerned about the maintainability of these things. It makes moving code around harder, and it forces somebody who has no interest in this debugging facility to thing about the categories. And naming them is hard even for somebody who does care. I'm not a fan of MEMSTALL_MEMDELAY, for example because it's way too non-descript. The distinction between MEMSTALL_WORKINGSET_REFAULT and MEMSTALL_WORKINGSET_THRASH is dubious as well. These are recipes for bit rot and making the code harder to hack on. I see two options to do this better: One is to use stack traces as identifiers instead of a made-up type. The other is to use the calling function as the id (see how kmalloc_track_caller() utilizes _RET_IP_). bpftrace can use the stack as a map key. So this should already work without any kernel modifications, using @start[tid, kstack]?