Implement parallel ARC eviction

[allanjude]

Sponsored-by: Expensify, Inc.
Sponsored-by: Klara, Inc.

Motivation and Context

Read and write performance can become limited by the arc_evict process being single threaded.
Additional data cannot be added to the ARC until sufficient existing data is evicted.

On many-core systems with TBs of RAM, a single thread becomes a significant bottleneck.

With the change we see a 25% increase in read and write throughput

Description

Use a new taskq to run multiple multiple arc_evict() threads at once, each given a fraction of the desired memory to reclaim

How Has This Been Tested?

Benchmarking with a full ARC to measure the performance difference.

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Performance enhancement (non-breaking change which improves efficiency)
• Code cleanup (non-breaking change which makes code smaller or more readable)
• Breaking change (fix or feature that would cause existing functionality to change)
• Library ABI change (libzfs, libzfs_core, libnvpair, libuutil and libzfsbootenv)
• Documentation (a change to man pages or other documentation)

Checklist:

• My code follows the OpenZFS code style requirements.
• I have updated the documentation accordingly.
• I have read the contributing document.
• I have added tests to cover my changes.
• I have run the ZFS Test Suite with this change applied.
• All commit messages are properly formatted and contain Signed-off-by.

[adamdmoss]

I've been casually testing this out (combined with the parallel_dbuf_evict PR) over the last couple of weeks (most recently, 5b070d1 ).

I've not been hammering it hard or specifically, just letting it do its thing with my messing-around desktop system.

Hit a probable regression today, though: while mv'ing a meager 8GB of files from one pool to another, all my zfs IO got really high-latency, and an iotop showed that the copy part of the move (this being a mv across pools, so in reality it's a copy-and-remove) was running at a painful few 100KB/sec, and the zfs arc_evict thread was taking a whole core... but just one core.

In time it all cleared up and of course I can't conclusively blame this PR's changes, but I left with two fuzzy observations:

• In many years of mucking around with ZFS I've never(?) seemed to get the 'arc_evict is pegging CPU badly' issue until I started testing this PR's changes (though I'm aware the issue occurs in the wild for folks on master/release ZFSes)
• arc_evict was only using one core as far as I can tell, so I guess the parallelism which is the point of this PR just wasn't kicking-in for some reason anyway and/or the spinning was happening outside of the parallelized part

[0mp]

I have updated the patch with a different logic for picking the default maximum number of ARC eviction threads. The new logic aims to pick the number that is one-eighth of the available CPUs, with a minimum of 2 and a maximum of 16.

[amotin]

one-eighth of the available CPUs, with a minimum of 2 and a maximum of 16.

Why would we need two evict threads on a single-core system? In that case I would probably prefer to disable taskqs completely. If that is a way to make it more logarithmic, then I would think about highbit(), though then it will grow pretty slow for very large systems, so that the limit of 16 will never be reached. But I am not exactly sure the faster growth would make sense, since it may cause more lock contentions in memory allocator, etc.

[allanjude]

one-eighth of the available CPUs, with a minimum of 2 and a maximum of 16.

Why would we need two evict threads on a single-core system? In that case I would probably prefer to disable taskqs completely. If that is a way to make it more logarithmic, then I would think about highbit(), though then it will grow pretty slow for very large systems, so that the limit of 16 will never be reached. But I am not exactly sure the faster growth would make sense, since it may cause more lock contentions in memory allocator, etc.

Right now, this is only enabled by a separate tunable, to enable multiple threads. So for the single CPU case, we don't expect it to be enabled. But for something like 4-12 core systems, we would want it to use at least 2 threads, and then grow from there, reaching 16 threads at 128 cores.

[amotin]

Right now, this is only enabled by a separate tunable, to enable multiple threads. So for the single CPU case, we don't expect it to be enabled.

Now that you mentioned it, I've noticed its been disabled by default. I don't like the idea to tune it manually in production depending on system size. I would prefer to to have reasonable automatic defaults.

[0mp]

Hey! So, here's what changed in the patch:

Formula

There is now a different formula for automatically scaling the number of evict threads when the parameter is set to 0. The formula is:

MIN(MAX(max_ncpus > 6 ? 2 : 1, ilog2(max_ncpus) + (max_ncpus >> 6)), 16);

It looks like this (the x axis is the CPU count and the y axis is the evict thread count):

Here's also a table:

CPUs	zfs_arc_evict_threads	Evict threads count	Using taskq?
1	0	1 (autoscaled)	No
2	0	1 (autoscaled)	No
5	0	1 (autoscaled)	No
6	0	2 (autoscaled)	Yes
1024	0	16 (autoscaled)	Yes
(not using autoscaling, CPU count is irrelevant)	1	1	No
(not using autoscaling, CPU count is irrelevant)	32	32	Yes

Less parameters

zfs_arc_evict_threads is now the only parameter exposed to control the evict thread count. The zfs_arc_evict_threads_parallel has been removed in favor of enabling the use of taskqs when there are two or more evict threads.

This approach has been suggested by @tonyhutter in another PR (#16487 (comment)).

Stability improvements

It is no longer possible to modify the actual evict threads count during runtime. Since the evict taskqs are only created during arc_init(), the module saves the actual number of evict threads it is going to use and does not care about changes to zfs_arc_evict_threads from then on. This behavior has been documented in the manual page.

[amotin]

Thanks for automating it. Few comments to that part, and please take a look on my earlier comments.

[amotin]

I am not sure it is right, but it seems GCC does no like it:

  module/zfs/arc.c: In function 'arc_evict_state':
  module/zfs/arc.c:4095:15: error: 'evarg' may be used uninitialized in this function [-Werror=maybe-uninitialized]
    evict_arg_t *evarg;
                 ^~~~~

[adamdmoss]

I gave this another spin (not in isolation though FYI - it was along with the parallel dbuf eviction PR) and got a repeat of the previously noted behavior. Seems to not be coincidence.

In stress-testing the intended use-case (chugging through data when the arc is already full) this PR seems benign and probably even beneficial - multiple arc reaper threads are active and busy, and throughput is very healthy.

However, later just puttering around in desktop usage under quite light reads I noticed that a reading app is blocked for several seconds at a time and the experience was quite unpleasant. Lo and behold, one or more arc_evict threads were spinning hard, eventually settling down but then spinning hard again for several seconds under further light read load. This pattern repeated until I rebooted (into a zfs branch minus this PR :) ).

[amotin]

Thanks. We are getting closer. ;)

[alex-stetsenko]

Thanks. We are getting closer. ;)

Is anything left unresolved?

[amotin]

@alex-stetsenko I'll take another look a bit later, but meanwhile would be good to fix style issue (some line is too long), squash it all into one commit and rebase on top of master.

[amotin]

We could limit nthreads to num_sublists. Multiple threads per sublist might not have much sense. It should normally be true, but zfs_multilist_num_sublists if configurable.

[alex-stetsenko]

I think it is worth checking num_sublist when zfs_arc_evict_threads_max is computed. Updated.

[amotin]

While it may be some better in practice by not creating extra threads, it dives into internal multilist details, which is not good. I wont insist, but I would not do it this way myself.

[amotin]

Couple more comments, but more important please take a look why no CI tests are passing. It seems they all time out for some reason. And since I saw it at least on several pushes of this PR but no others, it makes me think it is not a CI glitch, but something is wrong here.

[amotin]

While it may be some better in practice by not creating extra threads, it dives into internal multilist details, which is not good. I wont insist, but I would not do it this way myself.

[amotin]

@alex-stetsenko The CI still looks very unhappy, but now I have no guesses why.