Comment by marginalia_nu
5 days ago
So I have a buffer pool with O_DIRECT reads.
I implement read-ahead in the application by (optionally) preadv:ing a single read into multiple destination buffers in the pool, leaving them unpinned, since as long as you aren't up against the bandwidth limit of the drive, a larger read is generally as fast as multiple smaller one on modern hardware.
I've tried doing this with io_uring as well, but found just eating the preadv syscall cost was faster.
> a larger read is generally as fast as multiple smaller one on modern hardware.
Not always if by modern you mean NVMe drives. One synchronous preadv() for 256 KiB gives the kernel/device one big request but 16 independent asynchronous 16 KiB reads can be serviced concurrently. So the latter gives the NVMe controller 16 operations it can schedule in parallel. So depending on the workload and hardware, offsets, filesystem and request sizes that can give you lower aggregate latency or higher throughput.
If you submit 16 contiguous read requests the system will just merge them into one large read request.
Modern SSDs tolerate moderate queue depths very well, but piling on the I/O queue also incurs tail latency jitter unless you're able to ensure the queue depth stays in the moderate range and never goes higher. All else being equal, fewer larger requests is better for I/O latency (though read amplification for the sake of reading more data obviously doesn't help anyone). Though in this scenario, we're mostly comparing the syscall overhead of a single preadv against io_uring bookkeeping for multiple preads, regardless of how you submit the reads they end up being the same operation.
Doesn’t the Linux kernel automatically merge requests for contiguous regions into a single request? I’m not sure 16 read I/Os submitted to io_uring for a contiguous 64kib region will behave differently at the disk level than a single 64kib request. There’s a little bit extra work the kernel has to do to merge the requests but that should be negligible.
In a poorly written micro benchmark test it’s possible the kernel will fail to coalesce all of them because your submissions aren’t visible all at once (ie it starts submitting requests and doesn’t have an opportunity to merge). Whether that actually is possible to happen requires digging a bit more into the Linux kernel source.
1 reply →
I think that’s only true on paper; in practice it’ll be true only when you have competing reads (at a thousand-foot view) and it might be possible to algorithmically bundle a portion thereof. It originally let SCSI controllers attached to spinning rust HDDs optimize physical manipulation of the disk heads to optimized queued reads of data in a “traveling salesman” sort of way, but modern nand flash can only internally read a full page at a time (which may be much greater than even the apparent physical sector size) anyway and with a strictly constant cost regardless of the “physical location” of the data on the non-existent platter. Old drives had optimization constraints like higher sequential read speeds at the outside of the platter (more bytes per physical rotation) and extremely pathological cases for data written to the innermost tracks of the platter. Individual requests were much finer-grained and the latency was much more varied, so a request from app/thread X for as little as 512 bytes from one location could be cheaply piggy-backed on an existing request from app/thread Y to read multiple megabytes from a physically proximate source that would otherwise have seriously delayed or starved the queued waiting read while the outstanding request was serviced.
In fact, one consistently sees higher bulk IO numbers when using physical media that has been formatted with a large sector size compared to the old 512 byte fixed emulated size. You’d routinely see lower latency and higher IOPs with 4kn (HDDs or SSDs) than you would with 512e disks, even with SCSI or AHCI controllers that featured similar pipelining support to today’s NVME controllers (or even if you place a spinning rust HDD behind NVMe today!).
Mostly true but as someone else pointed out it’s still better to issue a single large multi-page contiguous read than the same read broken down into separate requests because the I/O queue isn’t infinite. So both the Linux kernel and the SSD microcontroller still benefit from merging contiguous requests on their end. However, if you issue it correctly at the application level that’s still going to be better in that you’re never going to encounter a situation where you accidentally don’t get the desired coalescing.
1 reply →
> disk heads to optimized queued reads of data in a “traveling salesman” sort of way
Luckily the read heads only have one degree of freedom, so the "elevator algorithm" is sufficient: https://en.wikipedia.org/wiki/Elevator_algorithm
Do this with io_uring with the preadv syscall. It’ll be the same or faster (faster only if you can do something else while waiting for I/O or you can submit multiple requests simultaneously - a single io_uring will be basically identical)
io_uring is significantly faster for certain workloads, but you can't expect that merely running a syscall via io_uring will somehow magically make it faster.
Replacing single sycalls with their equivalents in io_uring is generally slower than just making the syscall directly. io_uring still uses syscalls after all.
io_uring generally only wins if you can amortize its overhead across multiple simultaneous operations. Implementing readahead would be such a case, except you can accomplish the same amortization with a single preadv instead, which again turns into a single syscall for multiple reads.
That’s not the only case. A read call makes your thread unable to do anything for the duration of the read. Io_uring lets that same thread continue to handle other requests which themselves might generate more I/O that gets amortized.
The fair comparison isn’t 1 syscall on a single thread processing 1 task against io_uring. That would be insane because you clearly don’t have any performance requirements in such a workload already.
The closest realistic equivalent would be using Tokio’s spawn_blocking to do that 1 syscall vs doing that syscall in io_uring. It’s probably still more efficient if your benchmark literally is the cost of 1 syscall at a time but not by as much and io_uring in poll mode doesn’t even enter the kernel so it can actually outperform the syscall offloaded to a background thread (even though yes under the hood it’s the same kernel code).
3 replies →