lkml.org 
[lkml]   [2019]   [Sep]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] dm writecache: skip writecache_wait for pmem mode


On Thu, 5 Sep 2019, Huaisheng HS1 Ye wrote:

> > -----Original Message-----
> > From: Mikulas Patocka <mpatocka@redhat.com>
> > Sent: Wednesday, September 4, 2019 11:36 PM
> > On Wed, 4 Sep 2019, Huaisheng HS1 Ye wrote:
> >
> > >
> > > Hi Mikulas,
> > >
> > > Thanks for your reply, I see what you mean, but I can't agree with you.
> > >
> > > For pmem mode, this code path (writecache_flush) is much more hot than
> > > SSD mode. Because in the code, the AUTOCOMMIT_BLOCKS_PMEM has been
> > > defined to 64, which means if more than 64 blocks have been inserted
> > > to cache device, also called uncommitted, writecache_flush would be called.
> > > Otherwise, there is a timer callback function will be called every
> > > 1000 milliseconds.
> > >
> > > #define AUTOCOMMIT_BLOCKS_SSD 65536
> > > #define AUTOCOMMIT_BLOCKS_PMEM 64
> > > #define AUTOCOMMIT_MSEC 1000
> > >
> > > So when dm-writecache running in working mode, there are continuous
> > > WRITE operations has been mapped to writecache_map, writecache_flush
> > > will be used much more often than SSD mode.
> > >
> > > Cheers,
> > > Huaisheng Ye
> >
> > So, you save one instruction cache line for every 64*4096 bytes written to
> > persistent memory.
> >
> > If you insist on it, I can acknowledge it, but I think it is really an
> > over-optimization.
> >
> > Acked-By: Mikulas Patocka <mpatocka@redhat.com>
> >
> > Mikulas
>
> Thanks for your Acked-by, I have learned so much from your code.
>
> And I have another question about the LRU.
>
> Current code only put the last written blocks into the front of list
> wc->lru, READ hit doesn't affect the position of block in wc->lru. That
> is to say, if a block has been written to cache device, even there would
> be a lot of READ operation for that block next but without WRITE hit,
> which still would flow to the end of wc->lru, and eventually it would be
> written back.
>
> I am not sure whether this behavior disobeys LRU principle or not. But
> if this situation above appears, that would lead to some HOT blocks
> (without WRITE hit) had been written back, even READ hit many times. Is
> it worth submitting patch to adjust the position of blocks when READ
> hit? Just a discussion, I want to know your design idea.
>
> Cheers,
> Huaisheng Ye

The dm-writecache target is supposed to optimize writes, not reads.
Normally, there won't be any reads following a write, because the data
would be stored in the cache in RAM.

Mikulas

\
 
 \ /
  Last update: 2019-09-18 12:25    [W:0.069 / U:0.052 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site