Revert "revert recent VirtualFile asyncification changes (#5291)"

[problame]

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).

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
  • https://github.com/neondatabase/cloud/issues/8351
• changing code to ensure forward-progress once cache slots have been acquired
  • page_cache: inline {,try_}lock_for_write into memorize_materialized_page #5480
  • page_cache: ensure forward progress on miss #5482
  • tbd: VirtualFile::open_with_options is wasteful under pressure #6065
• reducing the amount of runnable tokio tasks
  • limit imitate accesses concurrency, using same semaphore as compactions #5578
  • concurrency-limit initial logical size calculation [v2] #6000
• fix bugs that caused unnecessary concurrency induced by connection handlers
  • Safekeeper timeline is deactivated without persisting state #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 => Epic: per-tenant read path throttling #5899
• pageserver-initiated load shedding

Related Issues

This PR unblocks the introduction of tokio-epoll-uring for asynchronous disk IO (Epic).

[github-actions]

2238 tests run: 2153 passed, 0 failed, 85 skipped (full report)

---

Flaky tests (1)

Postgres 15

• test_statvfs_pressure_usage: debug

Code coverage (full report)

• functions: 54.7% (10108 of 18486 functions)
• lines: 81.6% (58037 of 71140 lines)

---

_{The comment gets automatically updated with the latest test results
16cb2cd at 2024-01-09T19:45:43.299Z :recycle:}