Epic: scalable async disk IO (tokio-epoll-uring) #4744
Assignees
Labels
Comments
|
Development happens in https://github.com/neondatabase/tokio-epoll-uring |
This was referenced Dec 1, 2023
problame
added a commit
that referenced
this issue
Dec 4, 2023
…6000) Problem ------- Before this PR, there was no concurrency limit on initial logical size computations. While logical size computations are lazy in theory, in practice (production), they happen in a short timeframe after restart. This means that on a PS with 20k tenants, we'd have up to 20k concurrent initial logical size calculation requests. This is self-inflicted needless overload. This hasn't been a problem so far because the `.await` points on the logical size calculation path never return `Pending`, hence we have a natural concurrency limit of the number of executor threads. But, as soon as we return `Pending` somewhere in the logical size calculation path, other concurrent tasks get scheduled by tokio. If these other tasks are also logical size calculations, they eventually pound on the same bottleneck. For example, in #5479, we want to switch the VirtualFile descriptor cache to a `tokio::sync::RwLock`, which makes us return `Pending`, and without measures like this patch, after PS restart, VirtualFile descriptor cache thrashes heavily for 2 hours until all the logical size calculations have been computed and the degree of concurrency / concurrent VirtualFile operations is down to regular levels. See the *Experiment* section below for details. <!-- Experiments (see below) show that plain #5479 causes heavy thrashing of the VirtualFile descriptor cache. The high degree of concurrency is too much for In the case of #5479 the VirtualFile descriptor cache size starts thrashing heavily. --> Background ---------- Before this PR, initial logical size calculation was spawned lazily on first call to `Timeline::get_current_logical_size()`. In practice (prod), the lazy calculation is triggered by `WalReceiverConnectionHandler` if the timeline is active according to storage broker, or by the first iteration of consumption metrics worker after restart (`MetricsCollection`). The spawns by walreceiver are high-priority because logical size is needed by Safekeepers (via walreceiver `PageserverFeedback`) to enforce the project logical size limit. The spawns by metrics collection are not on the user-critical path and hence low-priority. [^consumption_metrics_slo] [^consumption_metrics_slo]: We can't delay metrics collection indefintely because there are TBD internal SLOs tied to metrics collection happening in a timeline manner (neondatabase/cloud#7408). But let's ignore that in this issue. The ratio of walreceiver-initiated spawns vs consumption-metrics-initiated spawns can be reconstructed from logs (`spawning logical size computation from context of task kind {:?}"`). PR #5995 and #6018 adds metrics for this. First investigation of the ratio lead to the discovery that walreceiver spawns 75% of init logical size computations. That's because of two bugs: - In Safekeepers: #5993 - In interaction between Pageservers and Safekeepers: #5962 The safekeeper bug is likely primarily responsible but we don't have the data yet. The metrics will hopefully provide some insights. When assessing production-readiness of this PR, please assume that neither of these bugs are fixed yet. Changes In This PR ------------------ With this PR, initial logical size calculation is reworked as follows: First, all initial logical size calculation task_mgr tasks are started early, as part of timeline activation, and run a retry loop with long back-off until success. This removes the lazy computation; it was needless complexity because in practice, we compute all logical sizes anyways, because consumption metrics collects it. Second, within the initial logical size calculation task, each attempt queues behind the background loop concurrency limiter semaphore. This fixes the performance issue that we pointed out in the "Problem" section earlier. Third, there is a twist to queuing behind the background loop concurrency limiter semaphore. Logical size is needed by Safekeepers (via walreceiver `PageserverFeedback`) to enforce the project logical size limit. However, we currently do open walreceiver connections even before we have an exact logical size. That's bad, and I'll build on top of this PR to fix that (#5963). But, for the purposes of this PR, we don't want to introduce a regression, i.e., we don't want to provide an exact value later than before this PR. The solution is to introduce a priority-boosting mechanism (`GetLogicalSizePriority`), allowing callers of `Timeline::get_current_logical_size` to specify how urgently they need an exact value. The effect of specifying high urgency is that the initial logical size calculation task for the timeline will skip the concurrency limiting semaphore. This should yield effectively the same behavior as we had before this PR with lazy spawning. Last, the priority-boosting mechanism obsoletes the `init_order`'s grace period for initial logical size calculations. It's a separate commit to reduce the churn during review. We can drop that commit if people think it's too much churn, and commit it later once we know this PR here worked as intended. Experiment With #5479 --------------------- I validated this PR combined with #5479 to assess whether we're making forward progress towards asyncification. The setup is an `i3en.3xlarge` instance with 20k tenants, each with one timeline that has 9 layers. All tenants are inactive, i.e., not known to SKs nor storage broker. This means all initial logical size calculations are spawned by consumption metrics `MetricsCollection` task kind. The consumption metrics worker starts requesting logical sizes at low priority immediately after restart. This is achieved by deleting the consumption metrics cache file on disk before starting PS.[^consumption_metrics_cache_file] [^consumption_metrics_cache_file] Consumption metrics worker persists its interval across restarts to achieve persistent reporting intervals across PS restarts; delete the state file on disk to get predictable (and I believe worst-case in terms of concurrency during PS restart) behavior. Before this patch, all of these timelines would all do their initial logical size calculation in parallel, leading to extreme thrashing in page cache and virtual file cache. With this patch, the virtual file cache thrashing is reduced significantly (from 80k `open`-system-calls/second to ~500 `open`-system-calls/second during loading). ### Critique The obvious critique with above experiment is that there's no skipping of the semaphore, i.e., the priority-boosting aspect of this PR is not exercised. If even just 1% of our 20k tenants in the setup were active in SK/storage_broker, then 200 logical size calculations would skip the limiting semaphore immediately after restart and run concurrently. Further critique: given the two bugs wrt timeline inactive vs active state that were mentioned in the Background section, we could have 75% of our 20k tenants being (falsely) active on restart. So... (next section) This Doesn't Make Us Ready For Async VirtualFile ------------------------------------------------ This PR is a step towards asynchronous `VirtualFile`, aka, #5479 or even #4744. But it doesn't yet enable us to ship #5479. The reason is that this PR doesn't limit the amount of high-priority logical size computations. If there are many high-priority logical size calculations requested, we'll fall over like we did if #5479 is applied without this PR. And currently, at very least due to the bugs mentioned in the Background section, we run thousands of high-priority logical size calculations on PS startup in prod. So, at a minimum, we need to fix these bugs. Then we can ship #5479 and #4744, and things will likely be fine under normal operation. But in high-traffic situations, overload problems will still be more likely to happen, e.g., VirtualFile cache descriptor thrashing. The solution candidates for that are orthogonal to this PR though: * global concurrency limiting * per-tenant rate limiting => #5899 * load shedding * scaling bottleneck resources (fd cache size (neondatabase/cloud#8351), page cache size(neondatabase/cloud#8351), spread load across more PSes, etc) Conclusion ---------- Even with the remarks from in the previous section, we should merge this PR because: 1. it's an improvement over the status quo (esp. if the aforementioned bugs wrt timeline active / inactive are fixed) 2. it prepares the way for #6010 3. it gets us close to shipping #5479 and #4744
problame
added a commit
that referenced
this issue
Jan 11, 2024
) This reverts commit ab1f37e. Thereby fixes #5479 Updated Analysis ================ The problem with the original patch was that it, for the first time, exposed the `VirtualFile` code to tokio task concurrency instead of just thread-based concurrency. That caused the VirtualFile file descriptor cache to start thrashing, effectively grinding the system to a halt. Details ------- At the time of the original patch, we had a _lot_ of runnable tasks in the pageserver. The symptom that prompted the revert (now being reverted in this PR) is that our production systems fell into a valley of zero goodput, high CPU, and zero disk IOPS shortly after PS restart. We lay out the root cause for that behavior in this subsection. At the time, there was no concurrency limit on the number of concurrent initial logical size calculations. Initial size calculation was initiated for all timelines within the first 10 minutes as part of consumption metrics collection. On a PS with 20k timelines, we'd thus have 20k runnable tasks. Before the original patch, the `VirtualFile` code never returned `Poll::Pending`. That meant that once we entered it, the calling tokio task would not yield to the tokio executor until we were done performing the VirtualFile operation, i.e., doing a blocking IO system call. The original patch switched the VirtualFile file descriptor cache's synchronization primitives to those from `tokio::sync`. It did not change that we were doing synchronous IO system calls. And the cache had more slots than we have tokio executor threads. So, these primitives never actually needed to return `Poll::Pending`. But, the tokio scheduler makes tokio sync primitives return `Pending` *artificially*, as a mechanism for the scheduler to get back into control more often ([example](https://docs.rs/tokio/1.35.1/src/tokio/sync/batch_semaphore.rs.html#570)). So, the new reality was that VirtualFile calls could now yield to the tokio executor. Tokio would pick one of the other 19999 runnable tasks to run. These tasks were also using VirtualFile. So, we now had a lot more concurrency in that area of the code. The problem with more concurrency was that caches started thrashing, most notably the VirtualFile file descriptor cache: each time a task would be rescheduled, it would want to do its next VirtualFile operation. For that, it would first need to evict another (task's) VirtualFile fd from the cache to make room for its own fd. It would then do one VirtualFile operation before hitting an await point and yielding to the executor again. The executor would run the other 19999 tasks for fairness before circling back to the first task, which would find its fd evicted. The other cache that would theoretically be impacted in a similar way is the pageserver's `PageCache`. However, for initial logical size calculation, it seems much less relevant in experiments, likely because of the random access nature of initial logical size calculation. Fixes ===== We fixed the above problems by - raising VirtualFile cache sizes - neondatabase/cloud#8351 - changing code to ensure forward-progress once cache slots have been acquired - #5480 - #5482 - tbd: #6065 - reducing the amount of runnable tokio tasks - #5578 - #6000 - fix bugs that caused unnecessary concurrency induced by connection handlers - #5993 I manually verified that this PR doesn't negatively affect startup performance as follows: create a pageserver in production configuration, with 20k tenants/timelines, 9 tiny L0 layer files each; Start it, and observe ``` INFO Startup complete (368.009s since start) elapsed_ms=368009 ``` I further verified in that same setup that, when using `pagebench`'s getpage benchmark at as-fast-as-possible request rate against 5k of the 20k tenants, the achieved throughput is identical. The VirtualFile cache isn't thrashing in that case. Future Work =========== We will still exposed to the cache thrashing risk from outside factors, e.g., request concurrency is unbounded, and initial size calculation skips the concurrency limiter when we establish a walreceiver connection. Once we start thrashing, we will degrade non-gracefully, i.e., encounter a valley as was seen with the original patch. However, we have sufficient means to deal with that unlikely situation: 1. we have dashboards & metrics to monitor & alert on cache thrashing 2. we can react by scaling the bottleneck resources (cache size) or by manually shedding load through tenant relocation Potential systematic solutions are future work: * global concurrency limiting * per-tenant rate limiting => #5899 * pageserver-initiated load shedding Related Issues ============== This PR unblocks the introduction of tokio-epoll-uring for asynchronous disk IO ([Epic](#4744)).
jcsp
changed the title
|
|
Deployed to staging: done: https://github.com/neondatabase/aws/pull/932 Will PR the changes from my benchmarking |
problame
changed the title
|
Moved from #2975 (comment) by @jcsp This week:
|
This was referenced Feb 5, 2024
|
Monitoring staging with new metrics through this week, to get more insight on whether our limits on locked memory are going to be a problem in prod. Maybe prod next week. |
|
This week:
|
|
This week:
|
Couldn't happen last week, will happen mid this week.
Do it after we've switched prod over. write path
on-demand downloads
|
|
This week:
|
Implemented & benchmarked last week, shipping this week. This week
|
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
The async
get_value_reconstruct_dataproject (GH project, final commit with good summary) converted more of the pageserver code base toasync fns.We had to revert that final commit due to performance considerations.
The current hypothesis is that most of the
spawn_blocking'ed calls were completely CPU-bound and too short-lived (single-digit microsecond range).Why were they short-lived? The current hypothesis is that
Under that hypothesis, spawn_blocking has too much overhead (CPU time => latency) for work that takes single-digit nanoseconds. In fact, microbenchmarks suggest that the breakeven point is at ca 25us of work on our systems.
(More details: https://www.notion.so/neondatabase/Why-we-needed-to-revert-my-async-get_value_reconstruct_data-patch-or-what-I-learned-about-spawn_b-91f28c48b7314765bdeed6e8cb38fdce?pvs=4 )
We considered switching to a design with
spawn_blockingThe problem is that we'd
So, this epic sets out to explore how we can continue to mostly rely on the kernel-page-cache for our read IO on the
Timeline::getcode path, in a way that is more scalable than spawn_blocking.This epic is the sister-epic to
Timeline::gettoasync fn#4743High-Level
Impl
Tasks
Follow-Ups not part of this epic
metadatafile #6890The text was updated successfully, but these errors were encountered: