On Thu, Nov 25, 2021 at 10:18 PM Michal Hocko <mhocko@xxxxxxxx> wrote: > > On Thu 25-11-21 12:31:33, Hao Lee wrote: > > On Thu, Nov 25, 2021 at 11:01:02AM +0100, Michal Hocko wrote: > > > On Thu 25-11-21 08:02:38, Hao Lee wrote: > > > > On Thu, Nov 25, 2021 at 03:30:44AM +0000, Matthew Wilcox wrote: > > > > > On Thu, Nov 25, 2021 at 11:24:02AM +0800, Hao Lee wrote: > > > > > > On Thu, Nov 25, 2021 at 12:31 AM Michal Hocko <mhocko@xxxxxxxx> wrote: > > > > > > > We do batch currently so no single task should be > > > > > > > able to monopolize the cpu for too long. Why this is not sufficient? > > > > > > > > > > > > uncharge and unref indeed take advantage of the batch process, but > > > > > > del_from_lru needs more time to complete. Several tasks will contend > > > > > > spinlock in the loop if nr is very large. > > > > > > > > > > Is SWAP_CLUSTER_MAX too large? Or does your architecture's spinlock > > > > > implementation need to be fixed? > > > > > > > > > > > > > My testing server is x86_64 with 5.16-rc2. The spinlock should be normal. > > > > > > > > I think lock_batch is not the point. lock_batch only break spinning time > > > > into small parts, but it doesn't reduce spinning time. The thing may get > > > > worse if lock_batch is very small. > > > > > > > > Here is an example about two tasks contending spinlock. Let's assume each > > > > task need a total of 4 seconds in critical section to complete its work. > > > > > > > > Example1: > > > > > > > > lock_batch = x > > > > > > > > task A taskB > > > > hold 4s wait 4s > > > > hold 4s > > > > > > > > total waiting time is 4s. > > > > > > 4s holding time is _way_ too long and something that this path should > > > never really reach. We are talking about SWAP_CLUSTER_MAX worth of LRU > > > pages. Sure there might be a bunch of pages freed that are not on LRU > > > but those are only added to a list. So again what is the actual problem? > > > > > > > The measurement unit in my example may not be rigorous and may confuse you. > > What I mean is the batch processing can only gives each task fairness to > > compete for this spinlock, but it can't reduce the wasted cpu cycles during > > spinning waiting, no matter what the batch size is. > > Correct. But isn't that a nature of pretty much any spinlock based > contention? There is not really much to be done except for removing the > lock. Yes... > Batching helps to amortize the spinning for the actual useful work > so that the spinning is not predominant. Trylocking to reduce that > spinning can be helpful only if you _know_ that other useful work could > be done and that you do not dalay the locked work way to much. This is a > tricky balance to make. Exactly, thanks for explanation. > > > No matter what the > > lock_batch is set, the following perf report won't change much. Many cpu > > cycles are wasted on spinning. Other tasks running on the same cores will be > > delayed, which is unacceptable for our latency-critical jobs. > > Could you share more about requirements for those? Why is unmapping in > any of their hot paths which really require low latencies? Because as > long as unmapping requires a shared resource - like lru lock - then you > have a bottle necks. We deploy best-effort (BE) jobs (e.g. bigdata, machine learning) and latency-critical (LC) jobs (e.g. map navigation, payments services) on the same servers to improve resource utilization. The running time of BE jobs are very short, but its memory consumption is large, and these jobs will run periodically. The LC jobs are long-run services and are sensitive to delays because jitters may cause customer churn. If a batch of BE jobs are finished simultaneously, lots of memory are freed, and spinlock contentions happen. BE jobs don't care about these contentions, but contentions cause them to spend more time in kernel mode, and thus, LC jobs running on the same cpu cores will be delayed and jitters occur. (The kernel preemption is disabled on our servers, and we try not to separate LC/BE using cpuset in order to achieve "complete mixture deployment"). Then LC services people will complain about the poor service stability. This scenario has occurred several times, so we want to find a way to avoid it. > > > I'm trying to > > find if it's possible to reduce the delay and the contention , after all, > > 59.50% is too high. This is why I post the thoughtless `cond_resched()` > > approach. > > What is the base for that 59.5%? Also how representative this is for > your sensitive workload? Unfortunately, we haven't found an accurate relationship between the acceptable jitters and the perf percentage. A reference datum is indeed essential, and we will try to setup it. Thanks for your tips. > > > Here is the perf report when executing ./usemem -j 4096 -n 20 10g -s 5 > > > > + 59.50% usemem [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > > + 4.36% usemem [kernel.vmlinux] [k] check_preemption_disabled > > + 4.31% usemem [kernel.vmlinux] [k] free_pcppages_bulk > > + 3.11% usemem [kernel.vmlinux] [k] release_pages > > + 2.12% usemem [kernel.vmlinux] [k] __mod_memcg_lruvec_state > > + 2.02% usemem [kernel.vmlinux] [k] __list_del_entry_valid > > + 1.98% usemem [kernel.vmlinux] [k] __mod_node_page_state > > + 1.67% usemem [kernel.vmlinux] [k] unmap_page_range > > + 1.51% usemem [kernel.vmlinux] [k] __mod_zone_page_state > -- > Michal Hocko > SUSE Labs
