Thank you for the additional background and suggestions.
I will resend the second version.
On 2025/1/17 20:55, Daniel 【外部账号】P. Berrangé wrote:
On Thu, Jan 16, 2025 at 01:37:44PM +0100, Kevin Wolf wrote:
Am 13.12.2024 um 16:56 hat Daniel P. Berrangé geschrieben:
On Thu, Nov 28, 2024 at 06:51:20PM +0800, [email protected] wrote:
From: Guoyi Tu <[email protected]>
Currently, disk I/O encryption and decryption operations are performed
sequentially
in the main thread or IOthread. When the number of I/O requests increases,
this becomes a performance bottleneck.
To address this issue, this patch use thread pool to perform I/O encryption
and decryption in parallel, improving overall efficiency.
We already have support for parallel encryption through use of IO threads
since approximately this commit:
commit af206c284e4c1b17cdfb0f17e898b288c0fc1751
Author: Stefan Hajnoczi <[email protected]>
Date: Mon May 27 11:58:50 2024 -0400
block/crypto: create ciphers on demand
Ciphers are pre-allocated by qcrypto_block_init_cipher() depending on
the given number of threads. The -device
virtio-blk-pci,iothread-vq-mapping= feature allows users to assign
multiple IOThreads to a virtio-blk device, but the association between
the virtio-blk device and the block driver happens after the block
driver is already open.
When the number of threads given to qcrypto_block_init_cipher() is
smaller than the actual number of threads at runtime, the
block->n_free_ciphers > 0 assertion in qcrypto_block_pop_cipher() can
fail.
Get rid of qcrypto_block_init_cipher() n_thread's argument and allocate
ciphers on demand.
Say we have QEMU pinned to 4 host CPUs, and we've setup 4 IO threads
for the disk, then encryption can max out 4 host CPUs worth of resource.
This is a lot of "if"s. Even just that it requires explicit
configuration and doesn't work out of the box would be a strong point
for me why having something that works by default (like a thread pool)
is worth it.
You're assuming that it's even possible to setup 4 iothreads which share
the load evenly. That's not a given at all. The only device that can
even make use of more than one iothread is virtio-blk. (And if we're
looking at all devices that exist in QEMU, most devices can't even make
use of a single iothread!) But if you do have a virtio-blk device, then
that setup means one iothread per queue. In a Linux guest, if all I/O
comes from a single CPU, then it will use the same queue and we'll have
three idle iothreads and one that is overloaded.
So in order to achieve a similar effect with iothreads, you must be
using virtio-blk, you must explicitly configure four iothreads and four
mappings of virtqueues to iothreads, and you must run a workload in the
guest that performs I/O from four different threads running on different
CPUs.
There are certainly good use cases for iothreads, but with this number
of preconditions, I don't think we can assume that they make it
unnecessary to parallelise things in other ways, too.
Ok thanks for the background, that all makes sense as a justification.
Lets capture a summary of this in the commit message.
How is this new proposed way to use a thread pool going to do better
than that in an apples-to-apples comparison ? ie allow same number
of host CPUs for both. The fundamental limit is still the AES performance
of the host CPU(s) that you allow QEMU to execute work on. If the thread
pool is allowed to use 4 host CPUs, it shouldn't be significantly different
from allowing use of 4 host CPUs for I/O threads surely ?
Having multiple different ways to support parallel encryption is not
ideal. If there's something I/O threads can't do optimally right
now, is it practical to make them work better ?
The limitations inside the guest obviously can't be changed by QEMU.
We could in theory add iothread support to more devices, though if they
don't have a concept of multiple queues that could be processed by
different threads, it's pretty pointless (apart from working around
limitations in the backends like you're suggesting here).
And of course, the most interesting one would be solving the
out-of-the-box aspect. This is far from trivial, because the optimal
configuration really depends on your workload, and nothing on the host
can know automatically what will eventually run in the guest. So it will
be tough finding defaults that improve this case without also hurting
other cases.
Yes, understood, I was missing the impact of the guest usage model.
With regards,
Daniel
