Very Computer

Hi,

I would appreciate any help deciding how much hardware to order
for a linux-based fileserver we want for our company.
We want to serve files for about 60-100 employees using NFS and
Samba. About 40 of these will be software developers possibly
running multiple compilations simultaneously. The following
is the configuration we are considering:

Pentium III 800 MHz
256 MB Ram.
Raid 5
Mycelex controller.

Would this be sufficient, or do we need a beefier box?

Thanks for any help. Please post replies to this newsgroup.

George.

--

Sent via Deja.com http://www.deja.com/
Before you buy.

I can't answer this question specifically, but I can offer suggestions for
posing a better question;

You should specify any additional services the box will be running, so that
prospective answer-writers don't have to guess. For example, are your
developers using a shell account on the box and compiling code locally? Or
are they running compile tools on their desktop machines and using the file
server in question as an NFS repository? Will the box be running any other
services like netatalk or LDAP...?

  I agree with the "need more info" assesment. However, 40 developers
compiling on the same box? I will start by assuming you mean doing the
compilations on the server and not locally. If so, then you are not even
close! This configuration would cover about 4 users compiling on it.
This assumes no remote X usage as well. Given the configuration though,
it makes more sense that you are doing the compilations on the users
local machine and then transfering files using smb or nfs. This makes
more sense. But, if you are using code control tools on the fileserver,
then the config again probably comes up short. A good rule of thumb
(well, maybe not good, but safe) is one proposed by Sun when I asked
them about how big a server I should get given a fixed number of people
and fixed workload (fileserver basically). They said that as long as it
is only IO, then you could say 30 users per CPU. This number drops
quickly as you get into more CPU intensive applications such as X or
gcc. Given that this is strictly marketing hype, then you might
extrapolate to 50 people per CPU, and figure 256 MB of ram per CPU for a
robust fileserver. I would say for the simple fileserver case, then you
would want a slower dual CPU (consider the Xeon at 500Mhz, excellent for
this kind of thing) with 512 MB or more of ram. The raid is an excellent
choice and needs no extra thought. Good luck.
--
remove nospam and replace with putzin to reply to me!
Pete Buelow
-- True believer in any OS other than MS!

Quote:> I can't answer this question specifically, but I can offer suggestions
for
> posing a better question;

> You should specify any additional services the box will be running, so
that
> prospective answer-writers don't have to guess. For example, are your
> developers using a shell account on the box and compiling code

locally? Or

Hi,

Thanks for responding. I apologize for not being more clear in my
previous post. We are not doing any kind of application development
on the machine. All the compilations will be done on the developers'
desktops. Also, nobody will run any apps like emacs or x-windows
on this machine. This is only for file serving using protocols
Samba and NFS.

I should also mention that we are leaning towards a Raid-5 solution
with a good Raid controller and 10000RPM drives. We have talked
to a few companies. After listening to our needs, the various
salesmen have proposed host-configurations ranging from 1CPU/256MB Ram
to 4CPU/2Gig Ram. As you might imagine, we are somewhat confused
at present!

George.

--

Sent via Deja.com http://www.deja.com/
Before you buy.

That's better. For this configuration, fileserver only, just get a
single CPU, a good amount of ram, and then the best disk IO stuff you
can. The faster the SCSI solution, the better, and forget IDE
altogether. SCSI is less CPU intensive and while more expensive, is more
robust for this kind of thing. Good luck.
--
remove nospam and replace with putzin to reply to me!
Pete Buelow
-- True believer in any OS other than MS!

Get something in between.  2 CPU / 1 G RAM sounds good to me.  4 CPU
systems are way over-priced, IMHO.  How about multiple file servers?  If
you have 40 software developers working on the same directory, you have
more serious problems than server performances.   I hope you are getting
SCSI disks, right?

In our last episode (Tue, 31 Oct 2000 09:55:17 -0800),
the artist formerly known as Edward Lee said:

The things to have "multiple" of are, for this sort of application,
the _I/O_ controllers.

- Multiple NICs could be of value if the traffic would saturate the
  segment the server is on.

- Multiple SCSI controllers could similarly be of value to provide
  additional disk I/O bandwidth.

The salescritters proposing multiple CPUs are primarily promoting
their own _sales_; the application in question is not likely to
saturate one CPU let alone four of them.

To be sure, the more memory, the more can get cached, and thus the
less work the disk drives need to do.

The most likely bottlenecks are nonetheless likely to fall somewhere
between:
a) Disk device latency,
b) Disk maximum throughput,
c) Network throughput.

The main CPU(s) don't have any involvement with these bottlenecks, and
I'd be surprised to see CPU get anywhere _close_ in importance to
those bottlenecks.
--

<http://www.hex.net/~cbbrowne/lsf.html>
Rules of the Evil Overlord #131. "I will never place the key to a cell
just out of a prisoner's reach." <http://www.eviloverlord.com/>

Quote:> I would appreciate any help deciding how much hardware to order
> for a linux-based fileserver we want for our company.
> We want to serve files for about 60-100 employees using NFS and
> Samba. About 40 of these will be software developers possibly
> running multiple compilations simultaneously. The following
> is the configuration we are considering:

> Pentium III 800 MHz
> 256 MB Ram.
> Raid 5
> Mycelex controller.

What is the LAN backbone speed?
What are the client connection speeds?
How many NICs in the server?
What if the server goes down?
Do developers have local client compilation/test space?  They should.

The bottlenecks will be throughput between the server's NICs and
switches first, then RAID controller throughput.

For real word NFS under the best conditions: 100bT NIC will do ~9MB/s,
4-100bT NICs will do ~30MB/s, 1000bT NIC will do ~23MB/s, 2-1000bT NICs
will do ~40MB/s.  More usually gives diminishing returns.

Gigabit sounds faster but seems to hit a ~25MB/s wall in an active LAN
environment even with a good gig backbone.  Relatively expensive.

RAID-5 does not give spectacular performance regardless of controller,
and nothing near disk or bus theoretical hardware burst rates.  Invite
the argumentative to demonstrate in your actual environment.

Try a single CPU box, good motherboard, 2-100bT NICs to the switches,
256MB mem and some reasonable CPU like Thunderbird or Pentium III, 7-800
MHz.  You can easily add 1-2 more NICs and another 256MB if needed.
File servers with a good I/O mix don't usually need big memory or a fast
CPU, but they always need big network bandwidth.

Gigabit sounds faster but seems to hit a wall around 25MB/s even with a
good gig backbone.  Relatively expensive.

Try different NFS mount options, buy good NICs and switches, forget ECC
memory, redundant power supplies and anything that adds complexity or is
hard to get.  Keep spares inhouse.

Make a plan for what happens if the box serving your developers and 60
email/office apps fails.  Usually this means either splitting users by
functional groups between multiple boxes (with spares to cover), or
network attached storage ala Network Appliance.

Make a plan to monitor server network, CPU, memory and RAID throughput
so there's no guesswork about client load patterns and weak links.
Incrementally fix the weakest link, repeat.

--
timothymoore
   bigfoot
     com

Quote:> I would appreciate any help deciding how much hardware to order
> for a linux-based fileserver we want for our company.
> We want to serve files for about 60-100 employees using NFS and
> Samba. About 40 of these will be software developers possibly
> running multiple compilations simultaneously. The following
> is the configuration we are considering:

> Pentium III 800 MHz
> 256 MB Ram.
> Raid 5
> Mycelex controller.

What is the LAN backbone speed?
What are the client connection speeds?
How many NICs in the server?
What if the server goes down?
Do developers have local client compilation/test space?  They should.

The bottlenecks will be throughput between the server's NICs and
switches first, then RAID controller throughput.

For real word NFS under the best conditions: 100bT NIC will do ~9MB/s,
4-100bT NICs will do ~30MB/s, 1000bT NIC will do ~23MB/s, 2-1000bT NICs
will do ~40MB/s.  More usually gives diminishing returns.

Gigabit sounds faster but seems to hit a ~25MB/s wall in an active LAN
environment even with a good gig backbone.  Relatively expensive.

RAID-5 does not give spectacular performance regardless of controller,
and nothing near disk or bus theoretical hardware burst rates.  Invite
the argumentative to demonstrate in your actual environment.

Try a single CPU box, good motherboard, 2-100bT NICs to the switches,
256MB mem and some reasonable CPU like Thunderbird or Pentium III, 7-800
MHz.  You can easily add 1-2 more NICs and another 256MB if needed.
File servers with a good I/O mix don't usually need big memory or a fast
CPU, but they always need big network bandwidth.

Gigabit sounds faster but seems to hit a wall around 25MB/s even with a
good gig backbone.  Relatively expensive.

Try different NFS mount options, buy good NICs and switches, forget ECC
memory, redundant power supplies and anything that adds complexity or is
hard to get.  Keep spares inhouse.

Make a plan for what happens if the box serving your developers and 60
email/office apps fails.  Usually this means either splitting users by
functional groups between multiple boxes (with spares to cover), or
network attached storage ala Network Appliance.

Make a plan to monitor server network, CPU, memory and RAID throughput
so there's no guesswork about client load patterns and weak links.
Incrementally fix the weakest link, repeat.

--
timothymoore
   bigfoot
     com

Guys,

Thanks very much for the expert advice. Based on all the feedback,
we have decided to go with a 1CPU, 512MB Ram, Raid-5, SCSI setup
with 10000 RPM drives and 2 Ethernet cards. We have also ordered
a tape drive and set up a system of incremental backups. Furthermore, we
have also ordered a honking UPS!

Once again, thanks very much indeed!

George.

--

Sent via Deja.com http://www.deja.com/
Before you buy.