SMP with Linux

SMP with Linux

Post by Marcos H. Woehrma » Thu, 30 Oct 1997 04:00:00



I have a dual Pentium Pro system on which I've been running Linux 2.0.30
(RedHat release 4.2) for about a month in single processor mode.  I'm now
ready to enable SMP mode but am concerned that this will effect the
system reliability (it's being used as a www and ftp server, among other
things).

Does anyone have any experience with SMP Linux 2.0.30?  Are there newer
kernels that would be better?  Will we see an increase in performance?

marcos

--
Not ready reading .signature                          |    Marcos H. Woehrmann

 
 
 

SMP with Linux

Post by Ram Samudral » Fri, 31 Oct 1997 04:00:00



>  Our principle problems seem to be with NFS:  sometimes that
>acts screwey, and with the ethernet:  we have it operating at 100MHz
>but it won't do full duplex for some reason.  (No, not even with
>the latest driver.)
>  However, the machine is a very decent compile and simulation
>server.  The 256M does wonders, absolute wonders, for performance
>on compile and linking.
>  The machine seems stable except for the occasional NFS weirdnesses.

Ditto.  I just installed 2.0.30 on our SMP machines, and it's great,
except for some NFS problems, which occured only when the boxes were
on a 100 Base T ethernet.  Incidentally, the same machines on an OLDER
kernel with horrible configuration were much more stable in terms of
NFS. I suspect the problem may be due to the kernel version, rather
than NFS in Linux.

--Ram


                                \  /
                                 \/  Valine long
                                \/   and prosper.

 
 
 

SMP with Linux

Post by Richard Westerve » Fri, 31 Oct 1997 04:00:00


Marcos,

    I can't talk much about 2.0.30, but I've had 2.0.29 running on a dual
    pentium system for about 4 months now....  no crashes or any other
    problems encountered so far...  

                                      -Richard Westerveld


>I have a dual Pentium Pro system on which I've been running Linux 2.0.30
>(RedHat release 4.2) for about a month in single processor mode.  I'm now
>ready to enable SMP mode but am concerned that this will effect the
>system reliability (it's being used as a www and ftp server, among other
>things).
>Does anyone have any experience with SMP Linux 2.0.30?  Are there newer
>kernels that would be better?  Will we see an increase in performance?
>marcos
>--
>Not ready reading .signature                          |    Marcos H. Woehrmann


 
 
 

SMP with Linux

Post by Zbigniew Zyc » Fri, 31 Oct 1997 04:00:00



Quote:> Does anyone have any experience with SMP Linux 2.0.30?  Are there newer
> kernels that would be better?  Will we see an increase in performance?

I have dual Pentium (notPro) and using  2.0.31-pre10.
2.0.31 did problems to me (deadlock on proc 0, aieee inturrupt handling, etc).

2.0.31-pre10 looks more stable.
(ftp://ftp.kernel.org/pub/linux/kernel/testing)

System has 128MB RAM, 4 * 4GB SCSI HDD, and is using very heavily
as news server. (load avg. 5-8)

Zbigniew Zych
--

 
 
 

SMP with Linux

Post by Michael Meissne » Fri, 31 Oct 1997 04:00:00



Quote:> I have a dual Pentium Pro system on which I've been running Linux 2.0.30
> (RedHat release 4.2) for about a month in single processor mode.  I'm now
> ready to enable SMP mode but am concerned that this will effect the system
> reliability (it's being used as a www and ftp server, among other things).

> Does anyone have any experience with SMP Linux 2.0.30?  Are there newer
> kernels that would be better?  Will we see an increase in performance?

I've been running a dual processor Dell GXpro since about April, mostly for
compile/edit/test cycles of the compiler.  I usually run kernel of the week
(currently 2.1.60), but did run for quite awhile with 2.0.29 (and a little with
2.0.31 when it came out).  I've had no problems with SMP behavior, though my
workload tends to stress the system differently than your does.

If you want to selectively boot SMP and non-SMP kernels of the same revision,
you probably want to change the revision of one of them so that the appropriate
SMP/non-SMP modules will be loaded (I just changed 2.1.58 to 2.1.1058 when I
built 2.1.58 non-SMP for instance).

You might want to consider applying patches to output/display how much time is
spent in each CPU.  I use the 2.1.xx version of the patch, and run a patched
procmeter, which gives a graph of how each processor is used.

Here is the README file from the patch for 2.0.27 (which I believe installs ok
for 2.0.31):

--------------------------------------------------------------------------------
README file - Linux kernel 2.0.27 patch to add per-cpu usage information to
              /proc/stat
--------------------------------------------------------------------------------

This patch is provided in the hope it will be helpful, but with no warranty of
any kind.

--------------------------------------------------------------------------------

1) Purpose of this patch

Under the current versions of the Linux kernel, it is not possible to get
detailed cpu times for each processor of an SMP system (i.e. user/nice/system
/idle time spent on each processor).

The /proc/stat pseudo-file only provides global time information.

This patch creates new entries in /proc/stat so that time spend in various
system states appears for each processor.

I decided to make this patch when I compiled the 'xosview' utility and realized
that it could not display detailed cpu activity for each processor on an SMP
system.

2) Changes made in /proc/stat

Before the patch, a typical /proc/stat dump would look like :

cpu  7942 5410 2591 185661
disk 8768 0 0 0
disk_rio 7166 0 0 0
disk_wio 1602 0 0 0
disk_rblk 10486 0 0 0
disk_wblk 4206 0 0 0
page 19359 5003
swap 72 367
intr 256109 201604 3934 0 0 41763 1 2 0 0 8804 0 0 0 1 0 0
ctxt 214658
btime 848340657
processes 298

With the patch applied on a 2-cpu system, it would look like :

cpu  7942 5410 2591 371322
cpu0  4354 3182 1811 185661
cpu1  3588 2228 780 185661
disk 8768 0 0 0
disk_rio 7166 0 0 0
disk_wio 1602 0 0 0
disk_rblk 10486 0 0 0
disk_wblk 4206 0 0 0
page 19359 5003
swap 72 367
intr 256109 201604 3934 0 0 41763 1 2 0 0 8804 0 0 0 1 0 0
ctxt 214658
btime 848340657
processes 298

The first line still displays the total times for the system (the idle time is
multiplied by the number of cpus, I believe this field was incorrect in
2.0 kernels).

The following lines give the times for each cpu present in the system (n lines
for n processors). Each processor is identified by the string 'cpu<i>', i being
an integer between 0 and (smp_num_cpus-1).

3) The modified files

/usr/src/linux/fs/proc/array.c
/usr/src/linux/kernel/sched.c
/usr/src/linux/include/kernel_stat.h

To apply the patch, simply copy the smp_cpu_usage_patch.diff file to
/usr/src/linux, and do a 'patch <filename>.diff'.

4) Remarks

First of all, I developped this little patch on a SuperMicro P6D0F PentiumPro
mainboard equipped with only 1 processor. So it is not very useful for me
up to now (it will be as soon as I can afford another P6-200 ! :-}).
Therefore, I could not test this patch as much as I would like to.

I read in Linux-related newsgroups that the time spent on other processors than
cpu#0 is considered to be always 100% system (or 100% user, don't remember). I
suppose this problem will show up clearly with this patch :-}. Maybe I will
try to know more about that if I get a 2nd cpu.

There should be no major problem, however, except one thing: the changes in the
format of /proc/stat may break some programs that rely on the fact that, say,
the swap info are in the 8th line of that file. I believe that, for efficiency
reasons, many programs do not check the content of the first field to find the
correct /proc/stat line.

As for now, I've been using the patched version for about 2 months without any
problem. I compiled a modified version of 'xosview' that displays several cpu
bars, one for the total time and one for each cpu. Total and cpu0 show the
same thing on my 1-cpu system, so up to that point everything seems to be OK...

You can find the xosview patch at :

http://www-isia.cma.fr/~forissie/smp_kernel_patch/

If you have remarks about this patch you can e-mail me. I will appreciate
feedback.

---
Jerome Forissier
ISIA
Ecole des Mines de Paris - Sophia Antipolis - France

--
Michael Meissner, Cygnus Solutions (Massachusetts office)
4th floor, 955 Massachusetts Avenue, Cambridge, MA 02139, USA

 
 
 

SMP with Linux

Post by Yves EYRIE » Mon, 17 Nov 1997 04:00:00


Hello.

        I have got installed the slackware version in a computer like yours,
and it work fine. The system use the two processor to do all jobs
needed.

A+

 
 
 

1. Best NIC card -- LInux 2.4 SMP -> LInux 2.4 SMP

Folks,

We have some nice Dell servers that are dual Pentium III's and
10,000 RPM scsi 160 disks running the new RedHat 7.2 (kernel 2.4.9...),
but we feel a bit "guilty" about using $15 Netgear FA310TX ethernet
cards in them.

Saw some PC type reviews and they liked the FA310TX -- we aren't
interested
in encryption/remote management/....   just TCP performance.

Thanks.

John Murtari

2. Detecting pipe mode in a shell

3. SMP and Linux

4. ZOOM modems and DIP Problem

5. smp with Linux

6. network.conf and ethx_MASKLEN setting

7. SMP and Linux question

8. smp or die ??!?!

9. SMP under LINUX

10. Wyse 7000i SMP and linux version?

11. Dell PowerEdge SMP and Linux

12. SMP and Linux

13. Help enable SMP on Linux 2.0.30