在 2025/9/16 15:34, Jason Wang 写道:
On Mon, Sep 15, 2025 at 7:07 PM Sheng Zhao <[email protected]> wrote:
在 2025/9/15 16:21, Jason Wang 写道:
On Mon, Sep 15, 2025 at 3:34 PM <[email protected]> wrote:
From: Sheng Zhao <[email protected]>
The allocation granularity of bounce pages is PAGE_SIZE. This may cause
even small IO requests to occupy an entire bounce page exclusively.
This sounds more like an issue of the IOVA allocating that use the
wrong granular?
Sorry, the previous email has a slight formatting issue.
The granularity of the IOVA allocator is customized during the
initialization of the vduse domain, and this value is also modified in
this commit.
Ok, let's add this to the changelog.
Btw, do you have perf numbers to demonstrate the benefit?
Thanks
For arm64 64KB base pages, compared with fixed 4KB bounce pages, using
native pages is more likely to fill up the bounce buffer(default 64MB),
resulting in I/O performance bottlenecks.
I used QEMU vduse-blk as the backend for testing write performance.
Below are the fio test results:
| native | fixed-4k
----------+--------------+-------------
numjobs=2 | bw=44.4MiB/s | bw=47.0MiB/s
iodepth=4 | iops=90.9k | iops=96.1k
----------+--------------+-------------
numjobs=4 | bw=58.8MiB/s | bw=61.1MiB/s
iodepth=4 | iops=120.3k | iops=125.4k
----------+--------------+-------------
numjobs=8 | bw=64.0MiB/s | bw=74.7MiB/s
iodepth=8 | iops=131.1k | iops=153.1k
----------+--------------+-------------
numjobs=16| bw=69.8MiB/s | bw=92.7MiB/s
iodepth=8 | iops=143.0k | iops=190.0k
Thanks
Thanks
The
kind of memory waste will be more significant on arm64 with 64KB pages.
So, optimize it by using fixed 4KB bounce pages.
Signed-off-by: Sheng Zhao <[email protected]>
Thanks
