[parisc-linux] New 53c700 driver preformance evaluation

[parisc-linux] New 53c700 driver preformance evaluation

[Daniel Engstrom]

Hi,

When I got the new driver running I also did some more bonnie runs with the
following drivers: 
2.4.9-pa32 with the old driver and disconnects disbled
2.4.9-pa61 with the new driver and tagged-queueing disabled
HPUX 10.20

I also did one test with putc and getc changed to the unlocked versions
to see what impact the thread safety had. It was large.

I wonder if the lower speed for block I/O with the new driver is due to the
disabled disconnects?

/Daniel

            --------Sequential Output-------- ---Sequential Input--- --Random--
            -Per Char- --Block--- -Rewrite--- -Per Char- --Block---- --Seeks----
Machine MB K/sec  %CPU K/sec %CPU  K/sec %CPU K/sec %CPU K/sec %CPU  /sec  %CPU
pa32_1 100   313  99.7 3682  68.9  1162  37.5  253  94.9  2545  43.7  75.5  11.9
pa32_2 100   314 100.0 3650  68.3  1175  38.0  252  94.8  2548  43.5  74.1  12.1
pa32_3 100   314 100.0 3641  67.9  1185  38.1  252  94.7  2567  43.2  71.9  12.0

pa61_1 100   314 100.0 2839  51.6   971  32.0  252  95.1  2465  42.5  72.5  11.5
pa61_2 100   313  99.7 2871  53.8   996  33.3  252  94.6  2478  43.0  63.3  10.1
pa61_3 100   313 100.0 2865  52.3   998  32.5  251  94.3  2484  43.2  69.3  11.1

pa61ul 100  1373  97.8 2941  53.7   997  32.4  993  71.1  2444  42.3  67.2  11.4

hpux_1 100   597  98.3 2882  38.4  1530  31.0  405  96.9  3644  41.9  72.8  13.4
hpux_2 100   594  98.4 2930  36.6  1526  28.5  403  96.8  3625  39.3  72.8  13.3
hpux_3 100   599  98.4 2883  37.1  1541  29.9  407  97.2  3646  41.7  72.0  13.1
 

 -- 

Re: [parisc-linux] New 53c700 driver preformance evaluation

[Grant Grundler]

Daniel Engstrom wrote:
> Hi,
> 
> When I got the new driver running I also did some more bonnie runs with the
> following drivers: 
> 2.4.9-pa32 with the old driver and disconnects disbled
> 2.4.9-pa61 with the new driver and tagged-queueing disabled
> HPUX 10.20

Daniel,
thanks for running these tests...always interesting to see.

The difference between -pa32 and -pa61 are numerous.
Lots of bug fixes many new minor features working.
Really need to compare same -paXX with/without disconnects
enabled in order to compare that variable.

> I wonder if the lower speed for block I/O with the new driver is due to the
> disabled disconnects?

I don't think so.
In general, disconnects increase context switching/cpu utilization
but improve SCSI bus utilization for environments with lots of
devices and lots of IO "in flight". ie more throughput at the cost of
more latency. For a single threaded test, I would expect disconnects
*enabled* to have lower throughput than when disabled.

Also, the fact that -pa61 output is more in line with hpux suggest -pa32
could be lying about it's performance. ie the data isn't on media
like bonnie expects. Same might be true for hpux on input. dunno.
Need to troll and bonnie or other disk io perf mailing lists.

BTW, another HPUX (only on workstations) thing is to set WCE (write
cache enable) on the SCSI device. This allows the SCSI device to claim
a write has completed before the data is actually on media.
WCE + disabling disconnects is usally a good combo for perf since the
device can immediately report good status. The problem is a
SCSI bus reset or power failure will blow cached data away.
That's why HPUX servers don't permit WCE on.

hth,
grant

Re: [parisc-linux] New 53c700 driver preformance evaluation

[Daniel Engstrom]

Grant,

On 2001.10.22 19:52 Grant Grundler wrote:
> The difference between -pa32 and -pa61 are numerous.
> Lots of bug fixes many new minor features working.
> Really need to compare same -paXX with/without disconnects
> enabled in order to compare that variable.

I disabled disconnect with pa61 and got similar results as with disconnects
disabled with pa32, so most of the difference may be the disconnects.

> > I wonder if the lower speed for block I/O with the new driver is due to the
> > disabled disconnects?

I phrased this in a bad way, the old driver which is faster have disconnects
disabled.

> For a single threaded test, I would expect disconnects
> *enabled* to have lower throughput than when disabled.
Which is what I saw even if it was not very clearly expressed.
I got ~400k/s better block throughput with the old driver with disconnects
disabled (due to some driver problem) than with the new driver.
 
> Also, the fact that -pa61 output is more in line with hpux suggest -pa32
> could be lying about it's performance. ie the data isn't on media
> like bonnie expects. Same might be true for hpux on input. dunno.
> Need to troll and bonnie or other disk io perf mailing lists.
> 
> BTW, another HPUX (only on workstations) thing is to set WCE (write
> cache enable) on the SCSI device. 
Can this behaviur be controlled? I use 10.20.

Another thing which I want to bring up is locking. It seems out getc() performance is stuck at 314chars/sec while HPUX gives almost 600, could it be
that HPUX lock primitive is twice as fast as ours?

/Daniel

-- 

Re: [parisc-linux] New 53c700 driver preformance evaluation

[Grant Grundler]

Daniel Engstrom wrote:
> Which is what I saw even if it was not very clearly expressed.
> I got ~400k/s better block throughput with the old driver with disconnects
> disabled (due to some driver problem) than with the new driver.

ah ok. It wasn't obvious to me.

> > BTW, another HPUX (only on workstations) thing is to set WCE (write
> > cache enable) on the SCSI device. 
> Can this behaviur be controlled? I use 10.20.

yes. iirc, it via scsictl command.

# scsictl
Usage: scsictl [-akq] [-m mode[=value]]... [-c command]... device

But HPUX calls WCE something else: "Immediate Reporting".
Keep that in mind when reading the man page.

> Another thing which I want to bring up is locking. It seems out getc() perfor
>   mance is stuck at 314chars/sec while HPUX gives almost 600, could it be
> that HPUX lock primitive is twice as fast as ours?

I think the lock primitive is the same - it's "ldcw".
It's something in the libc/syscall path but I have no clue what.

grant