Re: [PATCH v3 1/1] virtio-blk: avoid preallocating big SGL for data

From: Stefan Hajnoczi <stefanha@redhat.com>
To: Max Gurtovoy <mgurtovoy@nvidia.com>
Cc: hch@infradead.org, mst@redhat.com,
	virtualization@lists.linux-foundation.org, kvm@vger.kernel.org,
	israelr@nvidia.com, nitzanc@nvidia.com, oren@nvidia.com,
	linux-block@vger.kernel.org, axboe@kernel.dk
Subject: Re: [PATCH v3 1/1] virtio-blk: avoid preallocating big SGL for data
Date: Mon, 6 Sep 2021 16:09:45 +0100	[thread overview]
Message-ID: <YTYvOetMHvocg9UZ@stefanha-x1.localdomain> (raw)
In-Reply-To: <20210901131434.31158-1-mgurtovoy@nvidia.com>

[-- Attachment #1: Type: text/plain, Size: 2325 bytes --]

On Wed, Sep 01, 2021 at 04:14:34PM +0300, Max Gurtovoy wrote:
> No need to pre-allocate a big buffer for the IO SGL anymore. If a device
> has lots of deep queues, preallocation for the sg list can consume
> substantial amounts of memory. For HW virtio-blk device, nr_hw_queues
> can be 64 or 128 and each queue's depth might be 128. This means the
> resulting preallocation for the data SGLs is big.
> 
> Switch to runtime allocation for SGL for lists longer than 2 entries.
> This is the approach used by NVMe drivers so it should be reasonable for
> virtio block as well. Runtime SGL allocation has always been the case
> for the legacy I/O path so this is nothing new.
> 
> The preallocated small SGL depends on SG_CHAIN so if the ARCH doesn't
> support SG_CHAIN, use only runtime allocation for the SGL.
> 
> Re-organize the setup of the IO request to fit the new sg chain
> mechanism.
> 
> No performance degradation was seen (fio libaio engine with 16 jobs and
> 128 iodepth):
> 
> IO size      IOPs Rand Read (before/after)         IOPs Rand Write (before/after)
> --------     ---------------------------------    ----------------------------------
> 512B          318K/316K                                    329K/325K
> 
> 4KB           323K/321K                                    353K/349K
> 
> 16KB          199K/208K                                    250K/275K
> 
> 128KB         36K/36.1K                                    39.2K/41.7K

I ran fio randread benchmarks with 4k, 16k, 64k, and 128k at iodepth 1,
8, and 64 on two vCPUs. The results look fine, there is no significant
regression.

iodepth=1 and iodepth=64 are very consistent. For some reason the
iodepth=8 has significant variance but I don't think it's the fault of
this patch.

Fio results and the Jupyter notebook export are available here (check
out benchmark.html to see the graphs):

https://gitlab.com/stefanha/virt-playbooks/-/tree/virtio-blk-sgl-allocation-benchmark/notebook

Guest:
- Fedora 34
- Linux v5.14
- 2 vCPUs (pinned), 4 GB RAM (single host NUMA node)
- 1 IOThread (pinned)
- virtio-blk aio=native,cache=none,format=raw
- QEMU 6.1.0

Host:
- RHEL 8.3
- Linux 4.18.0-240.22.1.el8_3.x86_64
- Intel(R) Xeon(R) Silver 4214 CPU @ 2.20GHz
- Intel Optane DC P4800X

Stefan

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]