Re: data-logging on top of linux-2.4.19p7-compound-speedup-2.patch

Re: data-logging on top of linux-2.4.19p7-compound-speedup-2.patch

[Dieter Nützel]

On Sunday 05 May 2002 16:24, you wrote:
> On Fri, 2002-05-03 at 20:19, Dieter Nützel wrote:
> > Hello Chris,
> >
> > I would try the great speedup with data-logging (02-data-logging-2.diff)
> > on top of linux-2.4.19p7-compound-speedup-2.patch.
>
> Cool, thanks for giving it a try.
>
> > 01 and 04 are mostly (complete?) there.
>
> Right 01 and 04 are all of the speedup patches.
>
> > As you know I have a complete SCSI system.
> > Do I need 03-barrier-ide-datalogging.diff, too?
>
> No, you don't need the ide barrier code.  It won't hurt though.
>
> > SunWave1 src/linux# patch -p1 -E -N <
> > ../patches/ReiserFS/data-logging/02-data-logging-2.diff
> > patching file fs/buffer.c
>
> Take out the compound-speedup-2 patch and just use the ones supplied
> with the data logging code.  The speedups are the same, it just has a
> slightly different implementation of the VFS part.

So here are my result.

Kernel is with all my stuff (O(1)-scheduler, vm_34, preemption+lock-break, all 
pending patches and Chris latest data logging stuff.

2.4.19-pre7-O1-rml-pre-pc-dn2
+
01-commit_super-6-comp.diff
02-beta-data-logging-2.diff
03-barrier-ide-datalogging.diff
04-delayed_commit-datalogging.diff

Chris, I had to apply some stuff by hand but it works without a glitch when I 
mount without data logging.

I only noted two warnings during compilation in fs/reiserfs/inode.c due to the 
#if 0 ... clause.

th and jbegin_count were unused

fs/reiserfs/inode.c

[-]
void reiserfs_write_inode (struct inode * inode, int do_sync) {
    struct reiserfs_transaction_handle th ;
    int jbegin_count = 1 ;

    if (inode->i_sb->s_flags & MS_RDONLY) {
        reiserfs_warning("clm-6005: writing inode %lu on readonly FS\n",
                          inode->i_ino) ;
        return ;
    }
    /* memory pressure can sometimes initiate write_inode calls with sync == 
1,
    ** these cases are just when the system needs ram, not when the
    ** inode needs to reach disk for safety, and they can safely be
    ** ignored because the altered inode has already been logged.
    */
    if (do_sync && !(current->flags & PF_MEMALLOC)) {
        lock_kernel() ;
        reiserfs_commit_for_inode(inode) ;
#if 0
        journal_begin(&th, inode->i_sb, jbegin_count) ;
        reiserfs_update_sd (&th, inode);
        journal_end_sync(&th, inode->i_sb, jbegin_count) ;
#endif
        unlock_kernel() ;
    }
}
[-]

I used dbench for my benchmarks but maybe this is the wrong test?

Fist "normal" mount with "notail" only to get some basic numbers:

/dev/sda8 IBM DDYS-T18350N

32 clients started
..............................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................
....................................................................................................+.........+........................+.............++++.+.+...........++..........+.++.......+.+....+++..+.+.++.+++...+...+.++++********************************
Throughput 45.8549 MB/sec (NB=57.3186 MB/sec  458.549 MBit/sec)
14.580u 50.560s 1:33.13 69.9%   0+0k 0+0io 939pf+0w

/dev/sdb6 IBM DDRS-34560D

32 clients started
...............................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................+...........................................................................++++...........................+.....................++.+..................................................................+........+..............................................................................................................................................................
.......................................+.................+.....+.+...........................................+...................++...........+...+.++.................+..........+..........+................+........+.+.++.+..+********************************
Throughput 32.6923 MB/sec (NB=40.8654 MB/sec  326.923 MBit/sec)
13.740u 48.630s 2:10.21 47.8%   0+0k 0+0io 939pf+0w


Below only my "slow" DDRS disk:

-o data=journal

32 clients started
..............................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................
...............+....................................+.....................+........+.................+..................+...........+...............+...............+.............+........+..+........+++++++++++++++++.+.+.....+********************************
Throughput 10.8205 MB/sec (NB=13.5256 MB/sec  108.205 MBit/sec)
13.340u 57.410s 6:31.39 18.0%   0+0k 0+0io 939pf+0w

slow like a crawl
low user CPU: 0-6%
mid sys CPU:10-20%
continuous write out: 30.000-40.000 Kilobytes Paged Out per second


-o data=journal,notail

32 clients started
....................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................++++++++++++.......................................................................................................+..........+....................+...........+......+...................................+...............................................+.......................................................................+..+..............................+............+........................................
.......................................................................................................................................................+...................+............+.......+....+.............+......+.+....+********************************
Throughput 11.0182 MB/sec (NB=13.7727 MB/sec  110.182 MBit/sec)
13.390u 56.290s 6:24.39 18.1%   0+0k 0+0io 939pf+0w

CPU most the same
~30.000 - 47.000


-o data=journal,barrier=none,notail

32 clients started
..............................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................
.................................................................................+...........+.........................+......+..............+........................+.........+.....+........+.....++..+++++++++++++++++++....++********************************
Throughput 11.1542 MB/sec (NB=13.9428 MB/sec  111.542 MBit/sec)
13.800u 56.800s 6:19.71 18.5%   0+0k 0+0io 939pf+0w


-o barrier=none,notail

32 clients started
...................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................+...................................................................................................................................................................++++++++++.+....+......................................................
........................................+......+...................................+..............+...............+................+.........+.......+.+................+.....................++............+.+.+.+........++....+********************************
Throughput 35.9382 MB/sec (NB=44.9227 MB/sec  359.382 MBit/sec)
14.230u 49.600s 1:58.57 53.8%   0+0k 0+0io 939pf+0w


-o notail

32 clients started
..................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................+..................++++.....................................................+++.....+.......+.+......+....................++.........................+......................
...........................................................................................................+......................................................................+....+...............+......+..++.++.+.+.+++++.+********************************
Throughput 34.2208 MB/sec (NB=42.776 MB/sec  342.208 MBit/sec)
14.110u 51.430s 2:04.45 52.6%   0+0k 0+0io 959pf+0w


write cache ON (!!!)
-o notail

32 clients started
.........................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................++............+...........+...................................................+...........................................+..............................................................+...................................................+.++.......+............
............+............+.+..++++..............+..................................................................................................+.+.+........................................+......+....+.....+...+.+.....++++********************************
Throughput 36.226 MB/sec (NB=45.2825 MB/sec  362.26 MBit/sec)
13.860u 50.420s 1:57.62 54.6%   0+0k 0+0io 959pf+0w

So data=journal is very slow for me.
What's wrong?

Thanks,
	Dieter

Re: data-logging on top of linux-2.4.19p7-compound-speedup-2.patch

[Chris Mason]

On Fri, 2002-05-17 at 16:04, Dieter Nützel wrote:

> So data=journal is very slow for me.
> What's wrong?

Nothing ;-)  data=journal writes every data block twice, opnce to the
log and once to the main disk.  Most of the time this makes things
roughly 1/2 as fast as it was before.

The only time data=journal helps is when you've got a very fsync or
O_SYNC heavy application.  Try running iozone with -o on both
data=journal and without it, data=journal is 2 or 3 times as fast.

This is because we can do the sync by writing to the log instead of
writing to the log and having to seek back to the main disk.

-chris

Re: data-logging on top of linux-2.4.19p7-compound-speedup-2.patch

[Dieter Nützel]

On Monday 20 May 2002 15:46, Chris Mason wrote:
> On Fri, 2002-05-17 at 16:04, Dieter Nützel wrote:
> > So data=journal is very slow for me.
> > What's wrong?
>
> Nothing ;-)  data=journal writes every data block twice, opnce to the
> log and once to the main disk.  Most of the time this makes things
> roughly 1/2 as fast as it was before.

That's good "news".

> The only time data=journal helps is when you've got a very fsync or
> O_SYNC heavy application.

I were under the impression that I overlooked something.

>  Try running iozone with -o on both
> data=journal and without it, data=journal is 2 or 3 times as fast.

iozone3_98? What else?

> This is because we can do the sync by writing to the log instead of
> writing to the log and having to seek back to the main disk.

OK, know I'm enlightened ;-)

-Dieter