Unified allocation throttling

[amotin]

Motivation and Context

Existing allocation throttling had a goal to improve write speed by allocating more data to vdevs that are able to write it faster. But in the process it completely broke the original mechanism, designed to balance vdev space usage. With severe vdev space use imbalance it is possible that some with higher use start growing fragmentation sooner than others and after getting full will stop any writes at all. Also after vdev addition it might take a very long time for pool to restore the balance, since the new vdev does not have any real preference, unless the old one is already much slower due to fragmentation. Also the old throttling was request-based, which was unpredictable with block sizes varying from 512B to 16MB, neither it made much sense in case of I/O aggregation, when its 32-100 requests could be aggregated into few, leaving device underutilized, submitting fewer and/or shorter requests, or in opposite try to queue up to 1.6GB of writes per device.

Description

This change presents a completely new throttling algorithm. Unlike the request-based old one, this one measures allocation queue in bytes. It makes possible to integrate with the reworked allocation quota (aliquot) mechanism, which is also byte-based. Unlike the original code, balancing the vdevs amounts of free space, this one balances their free/used space fractions. It should result in a lower and more uniform fragmentation in a long run.

This algorithm still allows to improve write speed by allocating more data to faster vdevs, but does it in more controllable way. On top of space-based allocation quota, it also calculates minimum queue depth that vdev is allowed to maintain, and respectively the amount of extra allocations it can receive if it appear faster. That amount is based on vdev's capacity and space usage, but also applied only when the pool is busy. This way the code can choose between faster writes when needed and better vdev balance when not, with the choice gradually reducing together with the free space.

This change also makes allocation queues per-class, allowing them to throttle independently and in parallel. Allocations that are bounced between classes due to allocation errors will be able to properly throttle in the new class. Allocations that should not be throttled (ZIL, gang, copies) are not, but may still follow the rotor and allocation quota mechanism of the class without disrupting it.

How Has This Been Tested?

Test 1: 2 SSDs with 128GB and 256GB capacity written at full speed

Up to ~25% of space usage of the smaller one the SSDs are writing at about the same maximum speed. After that smaller device is gradually getting throttled to balance space usage. To the full space usage devices come with only few percent difference. Since users are typically discouraged to run at full capacity to reduce fragmentation, the performance at the beginning is more important than at the end.

Test 2: 2 SSDs with 128GB and 256GB capacity written at slower speed

Since we do not need more speed, the vdevs are keeping almost perfect space usage balance.

Test 3: SSD and HDD vdevs of the same capacity, but very different performance, written at full speed

While empty, the SSD is allowed to write 2.5 times faster than the HDD. With its space usage increase to ~50% the SSD is getting throttled to the HDD speed and after that even slower. To the full space usage devices come with only few percent difference.

Test 4: SSD and HDD vdevs of the same capacity, but very different performance, written at slower speed

Since we do not need more speed, the SSD is throttled to the HDD speed, but instead they are keeping almost perfect space usage balance.

Test 5: Second vdev addition

First the pool of one vdev is filled almost up to capacity. After that second vdev is added and the data is overwritten couple times. Single data overwrite is enough to re-balance the vdevs, even with some overshot, probably due to big sizes of TXG relative to device sizes used in the test and ZFS delayed frees.

Test 6: Parallel sequential write to 12x 5-wide RAIDZ1 of HDDs

block	old	new
16K	3076MiB/s	3621MiB/s
128K	3773MiB/s	4638MiB/s
1M	4434MiB/s	4541MiB/s
16M	4105MiB/s	4145MiB/s

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.

[amotin]

I am still thinking whether it would make sense to give smaller but fast vdev some smaller write boost at the beginning even on idle pool, so that it could be used more during reads, even by the cost of some lower write speeds later. Suppose it may have no a universal answer.

[allanjude]

This looks great, I especially like the concept of maintaining the minimum queue depth to keep all of the devices busy.

I know in general you are against adding tunables, but I wonder if a few of the magic numbers could be controllable.
Are there any kstats that we might want to add to be able to measure how this is working under load?

[amotin]

I know in general you are against adding tunables, but I wonder if a few of the magic numbers could be controllable.

I was thinking about some, but I've decided there is a pretty thin margin between different factors where the algorithm would work as planned, and it would be much easier to mess it up by random tuning rather than improve. I'll think about some that could have sense.

Are there any kstats that we might want to add to be able to measure how this is working under load?

I've used some custom tools to collect some data while benchmarking the code, but not sure what I would expose as kstat. I might think about it, but ideas are welcome.

[amotin]

@allanjude I've added metaslab_perf_bias tunable, controlling performance-based metaslab bias between never, on busy pool or always, to cover my above doubts.

[richardm1]

I'm curious about the impact of these changes to read performance; specifically when reading back data written through this new balance mechanism. Has this aspect been tested?

I mention this because I've been working to optimize sequential read performance of a small system with unbalanced (in both size and performance) rotating mirrors. I've noticed how the data "lands" on the platters can have a significant perf. impact when reading back said data later on.

On this box the most relevant and impactful knobs are metaslab_aliquot (lifting usually helps but only to a point) and metaslab_bias_enabled (disabling definitely helps) so a completely new write allocator is relevant to my interests. Write speed itself is far less important to me.

Is the code in a usable state where I (as a non-programmer who cannot even spell git) can kick the tires on it? I'd be thrilled to help with testing.

[amotin]

@richardm1 The mentioned module parameters are still applicable and working, and aside of me increasing metaslab_aliquot here to 2MB, considering slowly increasing linear density of modern HDDs, the on-disk placement should be similar, at least to what was designed originally, but later semi-broken. I don't have many read benchmarks for this, but you are welcome to try. The code should be totally usable to the best of my knowledge. I have no open issues about it.

[tonyhutter]

I'm admittedly not very familiar with the metaslab layer, but I don't see any more surface issues.

[pcd1193182]

In the future, for larger changes like this, it would be nice if during the review process follow-ups could be pushed as separate commits. This makes it possible to see what changed since the last review, which makes reviewing large diffs much easier. I understand that from the maintainers' side it's nice if everything is one commit at the end, but until that point having separate commits makes it significantly easier to track changes.

Other than the refcount issue, I think the change looks good.