Very Computer

Having spent a depressing few hours looking at all the latest Solaris 2.4,
bugs and patches, and broken patches....

Is it Sys V rel 4 thats slow and unreliable, or have Sun broken it?

In the multi-user benchmarks, Linux with a load of 16 Users was faster than
Solaris 2 with 4.   Presumably Sun can sell more MP machines, more memory, more ...

Ok, so Linux has less functionality, but what about all those much vaunted
performance improvements?   Sun says Solaris 2 is faster than SunOS 4, which
seems to be very dubious, considering the woes of those out there with 16MB RAM.

It's not funny anymore..    Does anyone acutally like Solaris2?

-- Rob

              And me with a pain in all the diodes down my left side

I'm not sure what you did or on what hardware.  We find Solaris 2.3/2.4
very reliable.  It also feels faster than SunOS 4.1.x.

Quote:>In the multi-user benchmarks, Linux with a load of 16 Users was faster than
>Solaris 2 with 4.   Presumably Sun can sell more MP machines, more memory, more ...

*the* benchmark?  Which benchmark was that?  And was it on the same
hardware?   Solaris 2.x handles a lot more things than linux does and
that has some disadvantage.  Almost all executables need to load
atleast 4 dynamic libraries, which is slow.  So the relative
overhead of starting an application in Solaris 2.x is huge.
But this overhead is required for the support of libintl and libw which
need the capability of dynamic linking.

Sun should and can bring this overhead down.

Solaris 2.x also has much more stuff in the kernel than Linux and
some of it is much faster/better than Linux (networking/NFS)

But I really wonder about these benchmarks.  A lot of people have
reported that Linux degrades very badly under load, which is something
we don't notice on Solaris 2.x.

Quote:>Ok, so Linux has less functionality, but what about all those much vaunted
>performance improvements?   Sun says Solaris 2 is faster than SunOS 4, which
>seems to be very dubious, considering the woes of those out there with 16MB RAM.

We don't have much problems with 16MB machines.  But you are confused.
Solaris 2.x does require more memory but it *is* faster with sufficient
memory (4-8MB more than SunOS 4.1.x)

Quote:>It's not funny anymore..    Does anyone acutally like Solaris2?

Yes.  It is much easier to maintain and configure than SunOS 4.1.x,
once you know SunOS 4.1.x as good as Solaris 2.x.

Casper

[....]

Quote:>In the multi-user benchmarks, Linux with a load of 16 Users was faster than
>Solaris 2 with 4.   Presumably Sun can sell more MP machines, more memory, more ...

[.....]

Quote:>It's not funny anymore..    Does anyone acutally like Solaris2?

Yes, people actually like Solaris 2. I have several machines that can
boot either Solaris x86 2.4 or Linux. I use both modes.....

On the same machine with identical hard drives, Linux fires up quicker
(if no e2fsck!) and seems a little peppier in response and in the file
system. Some of the things I do with Solaris involve wabi, answerbook and
the nice NFS server package.  I can guess that some of the apparent
overhead that Solaris  x86 seems to have with respect to Linux involve
the dynamic libraries and the extensive system info and statistics
gathering.

The Linux developers have done a fine job (in my opinion) and Sun has
done an excellent job of porting their pre-existing O/S to Intel
hardware. I give applause and virtual beers to both groups!

Kevin Martinez

--
------------------------------------------------------------------------
 Kevin Martinez                 |           Fry's Electronics: Where

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

Yes, far more than Linux.

Quote:>-- Rob
>              And me with a pain in all the diodes down my left side

David Holland
--
David Holland   "Of course, trusting the government with your privacy is like

                                                        -- John Perry Barlow

-> Having spent a depressing few hours looking at all the latest Solaris 2.4,
-> bugs and patches, and broken patches....
->
-> Is it Sys V rel 4 thats slow and unreliable, or have Sun broken it?
->
-> In the multi-user benchmarks, Linux with a load of 16 Users was faster than
-> Solaris 2 with 4.   Presumably Sun can sell more MP machines, more memory, more ...
->

I find SunOS / Solaris a dog.

For example, using a Sparc 1000E server with 1gb of ram, and two raid
5 disk arrays(fast and wide).  This system only gets 5megs per sec. xfer rate.
I pull the disk arrays, and put them on a SGI DM system (basic), and run the
same dd command (dd if=/dev/zero of=/array_mount_point/file.test bs=1024k
count=19531) which SGI gives about 14 megs per sec. xfer rate. So I called
Sun, and they thought that the 5megs per sec. was great!  Great?  I think
they need to have another look at the device drivers / kernel.

Now, my Linux box using the same test gets 8megs per sec.  It is a P5/90
system.

-> Ok, so Linux has less functionality, but what about all those much vaunted
-> performance improvements?   Sun says Solaris 2 is faster than SunOS 4, which
-> seems to be very dubious, considering the woes of those out there with 16MB RAM.

Don't let them lie to you.  That is a lie, and they know it.  Should you want
proof, ask almost anyone at Sun (Bay Area) who have been converted, if it (Solaris)
is faster, and they will tell you that Solaris sucks.  On power servers (like my
Sparc 1000E) to workstation's.

->
-> It's not funny anymore..    Does anyone acutally like Solaris2?
->
-> -- Rob
->
->               And me with a pain in all the diodes down my left side
->

--
-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-

-=  Unix System Admin.       |             flames=/dev/null  =-
-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-

Robert> Having spent a depressing few hours looking at all the latest Solaris
Robert> 2.4, bugs and patches, and broken patches....

Robert> Is it Sys V rel 4 thats slow and unreliable, or have Sun broken it?

Robert> In the  multi-user  benchmarks, Linux with   a load of  16  Users was
Robert> faster  than  Solaris  2 with 4.   Presumably  Sun  can sell  more MP
Robert> machines, more memory, more ...

Robert> Ok, so  Linux has less functionality, but  what  about all those much
Robert> vaunted performance improvements?  Sun says Solaris  2 is faster than
Robert> SunOS 4,  which seems  to be very   dubious, considering the  woes of
Robert> those out there with 16MB RAM.

  First of all  I have to say that  the guy from the  ix magazine has done  a
lousy job on testing Solaris 2.4 for x86.  All it showed was that 2.4 "as-is"
doesn't run very well  with just 16  MB. He  probably  not even though  about
adding more memory to see if this would change anything.  Solaris is known to
need a lot of memory.
  Just because the little info sheets that came with it said it would be able
to run Solaris 2.4 with 16 MB  doesn't mean it's enough for  a user load of 4
or even 16.
  The guy mentioned those info papers several times and complained they would
not  give enough information  and he had to  struggle several hours around to
figure out he needed an /etc/hostname.smc0 file for  his eithernet card.  But
somebody (in  de.comp.os.unix)  said that  the  hostname.smc0 file is  beeing
mentioned in  those installation docs.  So, either  this guy is  lying or the
one from the ix magazine didn't read it all through ...
  Installing Solaris just the  way it comes  down from the CD without ablying
atleast the recommented patches is silly. Not that it has to be that way, but
I guess he just didn't do that (atleast he  said nothing about it). Otherwise
I have no idea why he had so many crashes.

  But I was also surprised about the fact that Linux is that much faster than
Solaris. But is there any commercial unix  that would not loose against Linux
when it comes pure OS performance? (Again this is not an excuse of course.)
  And with every new  feature Linux becomes a  bit slower. So, wait till it's
able  of multi threading,   supports multiple cpus,  can  handle 16 bit  wide
characters, ...

  It is sad. Normally  the ix magazine gives  you really good information and
their test reports reflect the Real World (TM).

Robert> It's not funny anymore..    Does anyone acutally like Solaris2?

  Yes, atleast for Sparcs I do. SunOS  5.4 compared to  4.1.3 may be a little
bit faster, but all I can say is that it is faster than 5.3 and about as fast
as 4.1.3. But OpenWindows has   become a lot  faster than  under 4.1.3. So  I
would say Solaris 2.4 is a bit faster than 4.1.3.

Sven
--
LOAD "EMACS",8,1

I don't think you give us all the parameters in the test.
E.g., you can't even write 19GB files under Solaris 2.x.

On single disks under Solaris 2.x we get the max I/O bandwidth
available from the disk.   (6MB/s on F&W baracudas on writes,
9MB/s if you don't fsync)

The thruput on the disk array is probably limited by the speed of the
disks.  The disks Sun ships max out at perhaps 1-2MB/s.
So depending on how wide your stripes are, 5MB/s may well be the
limit of the disk.  In my experience, the limit of the disk
is the limit of Solaris 2.x, unless the processor is slow.

The SGIs aren't as fast as they look.  While they obviously do
I/O buffering faster, I don't find the thruput of normal
disks that enormous.  (you seem to be getting 14MB/s until
you fsync() the file, after which time the aggregrate thruputs drops
back to values expected from the disk thruput)
SGI does use a trick with page aligned buffers that aren't copied
to the kernel, which can be an enormous performance win.

Linux's filesystem implementation doens't use synchronous writes, and
I'm actually suprised that you only get 8MB/s on a P90, provided
all I/O fits in memory.  But we can't judge that as you obviously
didn't give us the test you really did run.

Quote:>Don't let them lie to you.  That is a lie, and they know it.  Should you want
>proof, ask almost anyone at Sun (Bay Area) who have been converted, if it (Solaris)
>is faster, and they will tell you that Solaris sucks.  On power servers (like my
>Sparc 1000E) to workstation's.

I'd be surprised if SunOS 4.1.x is faster on the SS1000E.  

We do find that the Solaris 2.4 desktop performance is considerably faster
than SunOS 4.1.x.  If you're asking the sysadmins, they may well dislike
the change for obvious reasons (having to relearn things)

Casper

: For example, using a Sparc 1000E server with 1gb of ram, and two raid
: 5 disk arrays(fast and wide).  This system only gets 5megs per sec. xfer rate.
: I pull the disk arrays, and put them on a SGI DM system (basic), and run the
: same dd command (dd if=/dev/zero of=/array_mount_point/file.test bs=1024k
: count=19531) which SGI gives about 14 megs per sec. xfer rate. So I called
: Sun, and they thought that the 5megs per sec. was great!  Great?  I think
: they need to have another look at the device drivers / kernel.
:
: Now, my Linux box using the same test gets 8megs per sec.  It is a P5/90
: system.

Thanks for the SGI plug :-)  I'll add to your fire.  By way of background,
I'm a kernel hacker that used to work at Sun and now works at SGI.  This
sort of I/O stuff is my area of expertise.  You can discount my ravings
because I work for SGI, but you should trust that I know what I'm talking
about in this area; ask people, even Sun engineers will admit it :-)

I can happily say that SGI blows the socks off of any Sun system in
terms of disk or network I/O.  I'm currently working on being able to
do at least 50MB/sec disk <-> net <-> remote machine.  I've done my own
testing on SGI systems and I can get 18MB/sec per fast&wide SCSI
string using just 3 disks.  Sure wish I had 100MB/sec SCSI strings...

On the lab nmachine we use for NFS performance, I've done 162MB/sec of
disk reads, sustained.  Personally.  I went into the lab and watched
all those pretty lights going.  I couldn't do any more than that because
I had 9 fast&wide SCSI strings.  Notice that the scaling was 100%
linear -- 162/9 == 18.

But that's no big deal, get a load of this.  SGI has a *file system*
that goes that fast.  Faster, in fact.  The xFS file system has been
measured running at 350MB/sec by Cray computer engineers (they're
bummin - their RAM disks aren't that fast, no kidding)  Internally,
we've had it up to 450MB/sec (I haven't seen that one myself, I don't
have that many disks).  By the way - the performance numbers are one
process doing normal reads.  No multi threading required.

In addition, SGI's network walks all over Sun's network.  I regularly get
over 70MB/sec through TCP/IP over HiPPI cables.  Try and do that with a
Sun (good luck).

SGI's MP's just blow the frigging doors off of Sun hardware &
software.  Sun may have more market share but they don't have a chance
of competing with SGI in terms of I/O.  SGI is at least a couple orders
of magnitude faster.  Sun doesn't stand a chance - if you care about
I/O, you buy an SGI.  

: -> It's not funny anymore..    Does anyone acutally like Solaris2?

No.  Never did, never will.  It's part of why I left Sun.
--
---

Quote:>SGI's MP's just blow the frigging doors off of Sun hardware &
>software.  Sun may have more market share but they don't have a chance
>of competing with SGI in terms of I/O.  SGI is at least a couple orders
>of magnitude faster.  Sun doesn't stand a chance - if you care about
>I/O, you buy an SGI.  

Wow 500mb per second off a single disk. It's a wonder I even come
to work the way you've got this all figured out.

---
=================================================================
| Kevin Clarke - Solaris Perf   |       "To err is human,       |

=================================================================

Quote:>Robert> Having spent a depressing few hours looking at all the latest Solaris
>Robert> 2.4, bugs and patches, and broken patches....

>Robert> Is it Sys V rel 4 thats slow and unreliable, or have Sun broken it?

>Robert> In the  multi-user  benchmarks, Linux with   a load of  16  Users was
>Robert> faster  than  Solaris  2 with 4.   Presumably  Sun  can sell  more MP
>Robert> machines, more memory, more ...

>Robert> Ok, so  Linux has less functionality, but  what  about all those much
>Robert> vaunted performance improvements?  Sun says Solaris  2 is faster than
>Robert> SunOS 4,  which seems  to be very   dubious, considering the  woes of
>Robert> those out there with 16MB RAM.

Solaris 2.4 for x86 is indeed a lot slower than other PC Unix systems.
I have run benchmarks on many different Unix versions, on many different
types of machines, and Solaris 2.4 (and also the previous 2.1 for that matter)
is indeed very slow.

I have also run a simulation of up to 15 simultaneous users, and below you
find some numbers that speek for themselves.
Note that when doing simple number crunching the same machine performs
the same with every Unix version. It's only when the kernel gets into heavy
action (switching processes, starting programs, piping, etc) that Solaris
becomes very slow.

On the machines below the AT 486-50 had 20MByte ram, the 486-66 had 32MByte
ram, (both with 256K cache) and the 386-25 had 8Mbyte ram (and 128K cache).

machine         unix-version    2       5       8       11      13      15
---------------------------------------------------------------------------
AT 486-50DX     Esix 4.0.4       1.0     2.5     3.8     5.3     6.0     7.1
Sparcstation 10 Sun OS 4.1.3     1.2     2.7     4.0     5.5     6.7     7.6
AT 486-50DX     AT&T sVr4.0  1.5     3.5     5.3     7.5     8.9    10.3
Mips 6850       Risc/OS 4.52     1.8     4.6     6.9     9.5    10.9    12.6
Sparcstation 2  Sun OS 4.1.1     2.1     4.9     7.9    10.4    12.2    13.9
Magnum 3000     Risc/OS 4.52     2.4     4.8     7.8    11.3    13.1    14.8
AT 386-25DX     System V-3.2     2.6     5.4     8.0    10.8    13.0    15.0
AT 486-66DX2    Solaris 2.1      2.4     5.9     9.3    13.0    15.1    17.3
AT 486-50DX     Solaris 2.4      2.8     5.9     9.5    12.8    15.1    17.4
DecStation 3100 Ultrix 3.1       2.8     6.5    10.2    13.9    16.3    18.6
AT 486-66DX2    Solaris 2.4      2.9     6.8    10.8    15.0    17.6    20.5
AT 286-12       SCO Xenix V2.2   7.7    19.0    30.5    42.0    49.6    57.5

These number lead to a very simple conclusion: for multiuser (multitasking)
Solaris reduces the speed of a PC-class machine to about 1/3 of the speed
the same machine achieves with other versions of Unix.

Or, in still other words: a 486-66 with Solaris is about as fast as a
386-25 was, 5 years ago.

This may be sad, but it is simply true.
I'm not saying Solaris is bad, because at least for me it seems to work well.
But if you want your machine to do a lot of things simultaneously, you should
probably look for another Unix version, because it's in the multitasking that
Solaris is at its worst.

Paul.

Quote:>machine             unix-version    2       5       8       11      13      15
>AT 286-12   SCO Xenix V2.2   7.7    19.0    30.5    42.0    49.6    57.5

    ^^^^^           ^^^^^

Yikes!  Why did you throw THAT in there?  Wouldn't SCO Unix 5.3.2.4.2 on
a 486 50 been a better benchmarker to compare to others?  Perhaps you just
had a 10 year old system lying around you wanted to try for the heck of it :)
--
  /--------------------------------------------------------------------------\
  | Mark A. Davis     | Lake Taylor Hospital | Norfolk,VA (804)-461-5001x431 |

  \--------------------------------------------------------------------------/

: >SGI's MP's just blow the frigging doors off of Sun hardware &
: >software.  Sun may have more market share but they don't have a chance
: >of competing with SGI in terms of I/O.  SGI is at least a couple orders
: >of magnitude faster.  Sun doesn't stand a chance - if you care about
: >I/O, you buy an SGI.  
:
: Wow 500mb per second off a single disk. It's a wonder I even come
: to work the way you've got this all figured out.

Jeez Kevin, you work on performance and you don't understand aggregate
I/O bandwidth?  Maybe you should stay home :-)  By the way, lowercase
"mb" normally means megabits and uppercase "MB" means megabytes.  We're
talking MB not mb.

Anyway, for those who care: nobody has 500MB disk drives.  SGI gets
their numbers by using (a) their logical volume manager to stripe the
drives, and (b) xFS, their high performance logging file system.
The interesting thing about SGI's hardware/software is that I've
yet to see non-linear scaling in I/O performance.  I striped 9
SCSI strings and it was exactly linear.  

And, given the independently verified performance numbers, I can
understand that Sun performance engineers would want to just give up...

Oh, yeah, did I mention 80MByte/sec through one networking interface
running standard TCP/IP?  Chuckle.  I like my new job, SGI seems to
care about being fastest.  
--
---