ping
On Wed 13 Feb 2019 05:48:53 PM CET, Alberto Garcia wrote:
> QEMU 2.12 (commit 1221fe6f636754ab5f2c1c87caa77633e9123622) introduced
> a new setting called l2-cache-entry-size that allows making entries on
> the qcow2 L2 cache smaller than the cluster size.
>
> I have been performing several tests with different cluster and entry
> sizes and all of them show that reducing the entry size (aka L2 slice)
> consistently improves I/O performance, notably during random I/O (all
> tests done with sequential I/O show similar results). This is to be
> expected because loading and evicting an L2 slice is more expensive
> the larger the slice is.
>
> Here are some numbers on fully populated 40GB qcow2 images. The
> rightmost column represents the maximum L2 cache size in both cases.
>
> Cluster size = 64 KB
> |-------------+--------------+--------------+--------------|
> | | 1MB L2 cache | 3MB L2 cache | 5MB L2 cache |
> |-------------+--------------+--------------+--------------|
> | 4KB slices | 6545 IOPS | 12045 IOPS | 55680 IOPS |
> | 16KB slices | 5177 IOPS | 9798 IOPS | 56278 IOPS |
> | 64KB slices | 2718 IOPS | 5326 IOPS | 57355 IOPS |
> |-------------+--------------+--------------+--------------|
>
> Cluster size = 256 KB
> |--------------+----------------+--------------+-----------------|
> | | 512KB L2 cache | 1MB L2 cache | 1280KB L2 cache |
> |--------------+----------------+--------------+-----------------|
> | 4KB slices | 8539 IOPS | 21071 IOPS | 55417 IOPS |
> | 64KB slices | 3598 IOPS | 9772 IOPS | 57687 IOPS |
> | 256KB slices | 1415 IOPS | 4120 IOPS | 58001 IOPS |
> |--------------+----------------+--------------+-----------------|
>
> As can be seen in the numbers, the only exception to the rule is when
> the cache is large enough to hold all L2 tables. This is also to be
> expected because in this case no cache entry is ever evicted so
> reducing its size doesn't bring any benefit.
>
> This patch sets the default L2 cache entry size to 4KB except when the
> cache is large enough for the whole disk.
>
> Signed-off-by: Alberto Garcia <be...@igalia.com>
