mm/zswap: change to writethrough

mm/zswap: change to writethrough

[Dan Streetman]

Seth, have you (or anyone else) considered making zswap a writethrough
cache instead of writeback?  I think that it would significantly help
the case where zswap fills up and starts writing back its oldest pages
to disc - all the decompression work would be avoided since zswap
could just evict old pages and forget about them, and it seems likely
that when zswap is full that's probably the worst time to add extra
work/delay, while adding extra disc IO (presumably using dma) before
zswap is full doesn't seem to me like it would have much impact,
except in the case where zswap isn't full but there is so little free
memory that new allocs are waiting on swap-out.

Besides the additional disc IO that obviously comes with making zswap
writethrough (additional only before zswap fills up), are there any
other disadvantages?  Is it a common situation for there to be no
memory left and get_free_page actively waiting on swap-out, but before
zswap fills up?

Making it writethrough also could open up other possible improvements,
like making the compression and storage of new swap-out pages async,
so the compression doesn't delay the write out to disc.

mm/zswap: change to writethrough

[Dan Streetman]

Seth, have you (or anyone else) considered making zswap a writethrough
cache instead of writeback?  I think that it would significantly help
the case where zswap fills up and starts writing back its oldest pages
to disc - all the decompression work would be avoided since zswap
could just evict old pages and forget about them, and it seems likely
that when zswap is full that's probably the worst time to add extra
work/delay, while adding extra disc IO (presumably using dma) before
zswap is full doesn't seem to me like it would have much impact,
except in the case where zswap isn't full but there is so little free
memory that new allocs are waiting on swap-out.

Besides the additional disc IO that obviously comes with making zswap
writethrough (additional only before zswap fills up), are there any
other disadvantages?  Is it a common situation for there to be no
memory left and get_free_page actively waiting on swap-out, but before
zswap fills up?

Making it writethrough also could open up other possible improvements,
like making the compression and storage of new swap-out pages async,
so the compression doesn't delay the write out to disc.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: mm/zswap: change to writethrough

[Bob Liu]

On 11/12/2013 03:12 AM, Dan Streetman wrote:
> Seth, have you (or anyone else) considered making zswap a writethrough
> cache instead of writeback?  I think that it would significantly help
> the case where zswap fills up and starts writing back its oldest pages
> to disc - all the decompression work would be avoided since zswap
> could just evict old pages and forget about them, and it seems likely
> that when zswap is full that's probably the worst time to add extra
> work/delay, while adding extra disc IO (presumably using dma) before
> zswap is full doesn't seem to me like it would have much impact,
> except in the case where zswap isn't full but there is so little free
> memory that new allocs are waiting on swap-out.
> 
> Besides the additional disc IO that obviously comes with making zswap
> writethrough (additional only before zswap fills up), are there any
> other disadvantages?  Is it a common situation for there to be no
> memory left and get_free_page actively waiting on swap-out, but before
> zswap fills up?
> 
> Making it writethrough also could open up other possible improvements,
> like making the compression and storage of new swap-out pages async,
> so the compression doesn't delay the write out to disc.
> 

I like this idea and those benefits, the only question I'm not sure is
would it be too complicate to implement this feature? It sounds like we
need to reimplement something like swapcache to handle zswap write through.

-- 
Regards,
-Bob

Re: mm/zswap: change to writethrough

[Bob Liu]

On 11/12/2013 03:12 AM, Dan Streetman wrote:
> Seth, have you (or anyone else) considered making zswap a writethrough
> cache instead of writeback?  I think that it would significantly help
> the case where zswap fills up and starts writing back its oldest pages
> to disc - all the decompression work would be avoided since zswap
> could just evict old pages and forget about them, and it seems likely
> that when zswap is full that's probably the worst time to add extra
> work/delay, while adding extra disc IO (presumably using dma) before
> zswap is full doesn't seem to me like it would have much impact,
> except in the case where zswap isn't full but there is so little free
> memory that new allocs are waiting on swap-out.
> 
> Besides the additional disc IO that obviously comes with making zswap
> writethrough (additional only before zswap fills up), are there any
> other disadvantages?  Is it a common situation for there to be no
> memory left and get_free_page actively waiting on swap-out, but before
> zswap fills up?
> 
> Making it writethrough also could open up other possible improvements,
> like making the compression and storage of new swap-out pages async,
> so the compression doesn't delay the write out to disc.
> 

I like this idea and those benefits, the only question I'm not sure is
would it be too complicate to implement this feature? It sounds like we
need to reimplement something like swapcache to handle zswap write through.

-- 
Regards,
-Bob

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: mm/zswap: change to writethrough

[Dan Streetman]

On Tue, Nov 12, 2013 at 4:11 AM, Bob Liu <bob.liu@oracle.com> wrote:
>
> On 11/12/2013 03:12 AM, Dan Streetman wrote:
>> Seth, have you (or anyone else) considered making zswap a writethrough
>> cache instead of writeback?  I think that it would significantly help
>> the case where zswap fills up and starts writing back its oldest pages
>> to disc - all the decompression work would be avoided since zswap
>> could just evict old pages and forget about them, and it seems likely
>> that when zswap is full that's probably the worst time to add extra
>> work/delay, while adding extra disc IO (presumably using dma) before
>> zswap is full doesn't seem to me like it would have much impact,
>> except in the case where zswap isn't full but there is so little free
>> memory that new allocs are waiting on swap-out.
>>
>> Besides the additional disc IO that obviously comes with making zswap
>> writethrough (additional only before zswap fills up), are there any
>> other disadvantages?  Is it a common situation for there to be no
>> memory left and get_free_page actively waiting on swap-out, but before
>> zswap fills up?
>>
>> Making it writethrough also could open up other possible improvements,
>> like making the compression and storage of new swap-out pages async,
>> so the compression doesn't delay the write out to disc.
>>
>
> I like this idea and those benefits, the only question I'm not sure is
> would it be too complicate to implement this feature? It sounds like we
> need to reimplement something like swapcache to handle zswap write through.

Simply converting to writethrough should be as easy as returning
non-zero from zswap_frontswap_store(), although
zswap_writeback_entry() also needs simplification to skip the
writeback.  I think it shouldn't be difficult; I'll start working on a
first pass of a patch.

Thanks!

Re: mm/zswap: change to writethrough

[Dan Streetman]

On Tue, Nov 12, 2013 at 4:11 AM, Bob Liu <bob.liu@oracle.com> wrote:
>
> On 11/12/2013 03:12 AM, Dan Streetman wrote:
>> Seth, have you (or anyone else) considered making zswap a writethrough
>> cache instead of writeback?  I think that it would significantly help
>> the case where zswap fills up and starts writing back its oldest pages
>> to disc - all the decompression work would be avoided since zswap
>> could just evict old pages and forget about them, and it seems likely
>> that when zswap is full that's probably the worst time to add extra
>> work/delay, while adding extra disc IO (presumably using dma) before
>> zswap is full doesn't seem to me like it would have much impact,
>> except in the case where zswap isn't full but there is so little free
>> memory that new allocs are waiting on swap-out.
>>
>> Besides the additional disc IO that obviously comes with making zswap
>> writethrough (additional only before zswap fills up), are there any
>> other disadvantages?  Is it a common situation for there to be no
>> memory left and get_free_page actively waiting on swap-out, but before
>> zswap fills up?
>>
>> Making it writethrough also could open up other possible improvements,
>> like making the compression and storage of new swap-out pages async,
>> so the compression doesn't delay the write out to disc.
>>
>
> I like this idea and those benefits, the only question I'm not sure is
> would it be too complicate to implement this feature? It sounds like we
> need to reimplement something like swapcache to handle zswap write through.

Simply converting to writethrough should be as easy as returning
non-zero from zswap_frontswap_store(), although
zswap_writeback_entry() also needs simplification to skip the
writeback.  I think it shouldn't be difficult; I'll start working on a
first pass of a patch.

Thanks!

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: mm/zswap: change to writethrough

[Seth Jennings]

On Mon, Nov 11, 2013 at 02:12:12PM -0500, Dan Streetman wrote:
> Seth, have you (or anyone else) considered making zswap a writethrough
> cache instead of writeback?  I think that it would significantly help
> the case where zswap fills up and starts writing back its oldest pages
> to disc - all the decompression work would be avoided since zswap
> could just evict old pages and forget about them, and it seems likely
> that when zswap is full that's probably the worst time to add extra
> work/delay, while adding extra disc IO (presumably using dma) before
> zswap is full doesn't seem to me like it would have much impact,
> except in the case where zswap isn't full but there is so little free
> memory that new allocs are waiting on swap-out.

There are two main benefits to zswap:
1) avoid write to swap space while the compressed pool is not full and
2) reduce application delay on anonymous page faults

Doing a writethough approach eliminates benefit 1 as there would
be just as much swap out activity with or without zswap.  However, you
still retain benefit 2.

The reclaim from zswap also becomes much simplier if the the compressed
page can be treated as clean.  It doesn't need to be written back and
can just be freed and the bit in the frontswap_map unset.

I'm not opposed to this.  The obstacle of writeback has influenced the
design a lot more than I would have liked.  You do lose the benefit of
decreased swap out traffic, but the real performance killer, application
latency on an anonymous page fault, is still greatly improved.

Seth

Re: mm/zswap: change to writethrough

[Seth Jennings]

On Mon, Nov 11, 2013 at 02:12:12PM -0500, Dan Streetman wrote:
> Seth, have you (or anyone else) considered making zswap a writethrough
> cache instead of writeback?  I think that it would significantly help
> the case where zswap fills up and starts writing back its oldest pages
> to disc - all the decompression work would be avoided since zswap
> could just evict old pages and forget about them, and it seems likely
> that when zswap is full that's probably the worst time to add extra
> work/delay, while adding extra disc IO (presumably using dma) before
> zswap is full doesn't seem to me like it would have much impact,
> except in the case where zswap isn't full but there is so little free
> memory that new allocs are waiting on swap-out.

There are two main benefits to zswap:
1) avoid write to swap space while the compressed pool is not full and
2) reduce application delay on anonymous page faults

Doing a writethough approach eliminates benefit 1 as there would
be just as much swap out activity with or without zswap.  However, you
still retain benefit 2.

The reclaim from zswap also becomes much simplier if the the compressed
page can be treated as clean.  It doesn't need to be written back and
can just be freed and the bit in the frontswap_map unset.

I'm not opposed to this.  The obstacle of writeback has influenced the
design a lot more than I would have liked.  You do lose the benefit of
decreased swap out traffic, but the real performance killer, application
latency on an anonymous page fault, is still greatly improved.

Seth

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>