[comp.dcom.lans] UN*X ffs vs. OS/2 hpfs for file server

norm@cfctech.cfc.com (Norman J. Meluch) (06/12/91)

We are currently trying to compare file servers here at CFC.

Right now we have a IBM PS/2 Model 95 (33MHz 486) running OS/2 1.3 EE
and Lan Manager 2.0 with a 3Com MC/32 (3C527) ethernet card.

There are those that suggest that we should try comparing this to an
AT&T Star Server E running UNIX SVR4.0.2.1 and their StarGroup software
version 3.4b, using a berkely fast file system partition.

Anyone have any feel as to whether these servers would be comparable
in terms of file access speed/throughput, or whether one is better
than another?

						- Norm.

-- 
Who: Norman J. Meluch - LAN/Unix Admin.       E-mail : norm@cfctech.cfc.com
Fax: (313)948-4975, Voice:(313)948-4809       Attmail: attmail!nmeluch
                                              Uucp   : ...!sharkey!cfctech!norm
"The other line moves faster.  -  Etorre's Observation"

kozowski@ohsu.edu (Eric Kozowski) (06/12/91)

In article <1991Jun11.175144.24736@cfctech.cfc.com> norm@cfctech.cfc.com (Norman J. Meluch) writes:
>Right now we have a IBM PS/2 Model 95 (33MHz 486) running OS/2 1.3 EE
>and Lan Manager 2.0 with a 3Com MC/32 (3C527) ethernet card.
>
>There are those that suggest that we should try comparing this to an
>AT&T Star Server E running UNIX SVR4.0.2.1 and their StarGroup software
>version 3.4b, using a berkely fast file system partition.
>
>Anyone have any feel as to whether these servers would be comparable
>in terms of file access speed/throughput, or whether one is better
>than another?

From the specs I've seen the AT&T StarServer E should blow the PS/2 95
out of the water.  I'm not familiar enough with OS/2 to say that ffs is
faster than the OS/2 file system (but I suspect that it is).

-- 
Eric Kozowski         
kozowski@ohsu.edu
Networks & Computing Dept.
Oregon Health Sciences University

jel@xerver.data.nokia.fi (Jerry Lahti) (06/14/91)

kozowski@ohsu.edu (Eric Kozowski) writes:
>From the specs I've seen the AT&T StarServer E should blow the PS/2 95
>out of the water.  I'm not familiar enough with OS/2 to say that ffs is
>faster than the OS/2 file system (but I suspect that it is).

Well, I would not bet on it.  As far as I know OS/2 HPFS uses quite 
similar techniques as the Berkeley FFS to optimize performance in
addition to some of its own ideas (like implementing directories
internally as B-trees).  In addition the HPFS-386 provided with
LAN Manager 2.0 integrates the network file server  very tightly
to the file system and disk drivers.  As a result it can often respond
to the most common file service requests from within the network
adapter interrupt handler.  This seems to provide quite a performance 
boost compared to running a LAN Manager server with the standard
OS/2 HPFS.  On the other hand, AT&T Stargroup implements the file
server as a user process and the required context switching is
quite likely hurt performance and especially server response times.

Jerry Lahti
Nokia Data Systems Oy, Networking/Network Operating Systems

pcg@aber.ac.uk (Piercarlo Grandi) (06/15/91)

On 11 Jun 91 17:51:44 GMT, norm@cfctech.cfc.com (Norman J. Meluch) said:

norm> Right now we have a IBM PS/2 Model 95 (33MHz 486) running OS/2 1.3 EE
norm> and Lan Manager 2.0 with a 3Com MC/32 (3C527) ethernet card.
norm> There are those that suggest that we should try comparing this to an
norm> AT&T Star Server E running UNIX SVR4.0.2.1 and their StarGroup software
norm> version 3.4b, using a berkely fast file system partition.

An incidental note: well, with SVR4 one should use the UFS/BSD FFS for
everything really.

I would like to oberve that maybe using either machine as a file server
is a bit funny; both machine can run one or two dozens users in
timesharing, without any problems. In particular the very large AT&T
StarServer (but please have a look into the larger Dell machines) can
run one or two dozens users with X windows, DOS, and Unix all together,
and probably deliver much better performance, and most importantly, many
less administrative headaches, than a bunch of PCs served by a OS/2. But
the PS/2 95 can run a lot of users too. Using a high speed machine with
high speed disks is wasted as a file server; in practice file service is
neither CPU bound nor IO bound, but network bound.

If you have a mainframe class machine (a 486@33 MHZ with suitable disks
is definitely mainframe class, running at 20MIPS and being capable of
running several GB of disk at several megabytes per second aggregate
thruput) you can just do a timesharing system. If the "mainframe" is
[34]86 based, you can easily run DOS under Unix on it (VP/IX, MERGE).

norm> Anyone have any feel as to whether these servers would be comparable
norm> in terms of file access speed/throughput, or whether one is better
norm> than another?

Purely as file server I think the OS/2 thing may be better, but I guess
that whatever is the difference it ought to be undetectable by users.
Both filesystems are very fast; the edge is with the BSD FFS, as the
OS/2 HPFS is not as well tuned. But a file server does not live or die
by disk IO performance, but by network overheads.

Networks of PC with a file server only make sense for economic reasons,
when you want to centralize your disk space to take advantage of
economies of scale, and any suitably slow machine will do. After all,
think, most PC based networks cannot physically transmit more than
1MB/sec. (or something of that order of magnitude) of data between all
stations. Having a server than can provide an aggregate thruput of more
than 1MB/sec. is entirely pointless, and any modern (16Mhz) 286 AT with
some decent ESDI or SCSI disk controller and a couple of disks will do.
--
Piercarlo Grandi                   | ARPA: pcg%uk.ac.aber@nsfnet-relay.ac.uk
Dept of CS, UCW Aberystwyth        | UUCP: ...!mcsun!ukc!aber-cs!pcg
Penglais, Aberystwyth SY23 3BZ, UK | INET: pcg@aber.ac.uk

les@chinet.chi.il.us (Leslie Mikesell) (06/18/91)

In article <PCG.91Jun14183441@aberdb.aber.ac.uk> pcg@aber.ac.uk (Piercarlo Grandi) writes:

>norm> Right now we have a IBM PS/2 Model 95 (33MHz 486) running OS/2 1.3 EE
>norm> and Lan Manager 2.0 with a 3Com MC/32 (3C527) ethernet card.
>norm> There are those that suggest that we should try comparing this to an
>norm> AT&T Star Server E running UNIX SVR4.0.2.1 and their StarGroup software
>norm> version 3.4b, using a berkely fast file system partition.

>I would like to oberve that maybe using either machine as a file server
>is a bit funny; both machine can run one or two dozens users in
>timesharing, without any problems.

I can't comment on the original question, but I would like to disagree
with most of your premises.

>Using a high speed machine with
>high speed disks is wasted as a file server; in practice file service is
>neither CPU bound nor IO bound, but network bound.

I've only rarely seen this on a 1M starlan network, I doubt if it would
be true on the 10M ethernet or 16M token ring net that you would be
likely to install with a 486 server.

>If you have a mainframe class machine (a 486@33 MHZ with suitable disks
>is definitely mainframe class, running at 20MIPS and being capable of
>running several GB of disk at several megabytes per second aggregate
>thruput) you can just do a timesharing system. If the "mainframe" is
>[34]86 based, you can easily run DOS under Unix on it (VP/IX, MERGE).

The 386 machines I use don't manage several megabytes/second through the
unix file system, but maybe that will improve with SysVr4.  Vp/ix is
not a general solution for multi-user dos access because it don't present
a network interface for network-aware programs and it typically consumes
about 10x the CPU of a file server handling a PC.

>Networks of PC with a file server only make sense for economic reasons,
>when you want to centralize your disk space to take advantage of
>economies of scale, and any suitably slow machine will do.

This is just not true.  The real reasons for PC networks are:
  (A) Dos programs can do everything a typical office user needs, and
      many have no affordable unix equivalent.  If you work with outside
      resources (and who doesn't) the odds are about 100 times better    
      that their programs and media are dos instead of your particular
      version of unix.  If your people travel with laptops, they have
      to be able to copy everything off to them.
  (B) It doesn't make any sense to have office users spend time installing
      their own software or making backups.  Nor should they be able
      to accidentally erase or modify the installed programs and data
      files that make the office work.
  (C) Everybody needs to be able to share data.

The economy of sharing disk space and resources like printers is really
trivial compared to centralizing the machine management duties. (University
people with cheap student labor may disagree...).

>After all,
>think, most PC based networks cannot physically transmit more than
>1MB/sec. (or something of that order of magnitude) of data between all
>stations. Having a server than can provide an aggregate thruput of more
>than 1MB/sec. is entirely pointless, and any modern (16Mhz) 286 AT with
>some decent ESDI or SCSI disk controller and a couple of disks will do.

Yes, if all you do is read and write from a few large files.  In practice,
that's not the case.  In a fair sized network you will have people making
and breaking connections all the time, the file server will provide
security functions and file and record locking with some associated overhead.
Most DOS programs these days provide internal directory reading functions
which may involve a great deal of overhead for the server to process.
Also, the server may be simultaneously handling many other functions like
mail transfers, gateways to mainframes, fax boards, and logged in users.
The server may also allow direct program execution with i/o piped to and
from the client.  The ability to perform these other functions transparently
along with the file services should be considered when choosing the server
platform.

Les Mikesell
  les@chinet.chi.il.us