Re: [PATCH 0/7] qcow2: compressed write cache

[Max Reitz]

On 29.01.21 17:50, Vladimir Sementsov-Ogievskiy wrote:
> Hi all!
>
> I know, I have several series waiting for a resend, but I had to switch
> to another task spawned from our customer's bug.
>
> Original problem: we use O_DIRECT for all vm images in our product, it's
> the policy. The only exclusion is backup target qcow2 image for
> compressed backup, because compressed backup is extremely slow with
> O_DIRECT (due to unaligned writes). Customer complains that backup
> produces a lot of pagecache.
>
> So we can either implement some internal cache or use fadvise somehow.
> Backup has several async workes, which writes simultaneously, so in both
> ways we have to track host cluster filling (before dropping the cache
> corresponding to the cluster).  So, if we have to track anyway, let's
> try to implement the cache.

I wanted to be excited here, because that sounds like it would be very 
easy to implement caching.  Like, just keep the cluster at 
free_byte_offset cached until the cluster it points to changes, then 
flush the cluster.

But then I see like 900 new lines of code, and I’m much less excited...

> Idea is simple: cache small unaligned write and flush the cluster when
> filled.
>
> Performance result is very good (results in a table is time of
> compressed backup of 1000M disk filled with ones in seconds):

“Filled with ones” really is an edge case, though.

> ---------------  -----------  -----------
>                   backup(old)  backup(new)
> ssd:hdd(direct)  3e+02        4.4
>                                  -99%
> ssd:hdd(cached)  5.7          5.4
>                                  -5%
> ---------------  -----------  -----------
>
> So, we have benefit even for cached mode! And the fastest thing is
> O_DIRECT with new implemented cache. So, I suggest to enable the new
> cache by default (which is done by the series).

First, I’m not sure how O_DIRECT really is relevant, because I don’t 
really see the point for writing compressed images.

Second, I find it a bit cheating if you say there is a huge improvement 
for the no-cache case, when actually, well, you just added a cache.  So 
the no-cache case just became faster because there is a cache now.

Well, I suppose I could follow that if O_DIRECT doesn’t make much sense 
for compressed images, qemu’s format drivers are free to introduce some 
caching (because technically the cache.direct option only applies to the 
protocol driver) for collecting compressed writes.  That conclusion 
makes both of my complaints kind of moot.

*shrug*

Third, what is the real-world impact on the page cache?  You described 
that that’s the reason why you need the cache in qemu, because otherwise 
the page cache is polluted too much.  How much is the difference really? 
 (I don’t know how good the compression ratio is for real-world images.)

Related to that, I remember a long time ago we had some discussion about 
letting qemu-img convert set a special cache mode for the target image 
that would make Linux drop everything before the last offset written 
(i.e., I suppose fadvise() with POSIX_FADV_SEQUENTIAL).  You discard 
that idea based on the fact that implementing a cache in qemu would be 
simple, but it isn’t, really.  What would the impact of 
POSIX_FADV_SEQUENTIAL be?  (One advantage of using that would be that we 
could reuse it for non-compressed images that are written by backup or 
qemu-img convert.)

(I don’t remember why that qemu-img discussion died back then.)


Fourth, regarding the code, would it be simpler if it were a pure write 
cache?  I.e., on read, everything is flushed, so we don’t have to deal 
with that.  I don’t think there are many valid cases where a compressed 
image is both written to and read from at the same time.  (Just asking, 
because I’d really want this code to be simpler.  I can imagine that 
reading from the cache is the least bit of complexity, but perhaps...)

Max