Re: Building Big Ass Linux Machine, what are the limits?

Gregory Maxwell (linker@z.ml.org)
Fri, 25 Sep 1998 10:16:20 -0400 (EDT)


On Thu, 24 Sep 1998, Jim McQuillan wrote:

> Hello all,
>
> I need to build a big machine to replace a
> Compaq Proliant running SCO OpenServer.

Okay.

> The system needs to run several Progress databases (12gb and growing)
> and I am
> looking for maximum disk throughput for about 150 users accessing via
> Telnet.

Alright.

> Some are telling me that I should be looking at big hardware, such
> as HP9000, SGI or Sun. The type of equipment that would cost upwards of
> $100k.

You prob would be better off with one of those.. Not because of Linux more
because of x86 hardware. But for the money....

> I want to try it with Linux.

Great.

> Here are my preferences:
>
> - Intel 450 Single or Dual, the application is really NOT cpu bound.
> I need Intel because until Progress releases a Linux version
> I will be using the SCO version with iBCS. And, as a safety net,
> if Linux isn't stable in this setup, I can put SCO on it
> (I really don't want to do that).

I'd get a single in a dual board. The dual boards tend to be of a higher
quality too.. If this is going to be occuring a few months into the future
(like when kernel 2.2 is out) then I'd say go with dual.

> - 512mb ram. Can I go bigger?, whats the max?
> Current system has 512mb and is only using about 300mb of it.

I think you'll find Linux to use less memory then SCO. Still, you should
be able to goto at least a gig in most high end boards today if you use
256meg dimms. Keep expandibility in mind. I highly recommend ECC memory as
there is almost nothing worse then trying to track down a intermient
memory problem that is silently corrupting memory.

> - 4 Adaptec 2940UW Host adaptors with 4-4gb Ultra wide disks on each.
> 16 drives total. Isn't there a limit on number of scsi disks?
> If there is a limit, I could change it to a total of 8 9gb drives.
> Progress has it's own Multi-volume support and can utilize as many
> drives as I can give it.
> I figure: more spindles + more heads = better disk performance.

Well there is a limit of 16 but you can hack it to increase it. Someone
else here should have more info.

About Progress' multi volume support? Does it stripe your data across
volumes?

Do you realize that you have NO reduncy and if a drive flakes out, it all
goes up in smoke?

Do you need 64gigs of storage? or would 32 do? I would recommend putting
those 16 drives (use 10,000 or at least 7,200 rpm ones) on a nice
multichannel raid controller (I've used DPT if you use them get the high
end model, I've also heard ICP Vortex is nice).... Set it up in a Raid-0/1
array. This gives you the best possible performance, and reduncy at the
price of 50% of your storage. (but it's fast as hell)

> Is Linux ready for the Enterprise Server market?

I feel that it is. I've implimented some really big Linux instlations and
I've certantly had no more problems with Linux then a simmlar sized novell
instlation. (and the Linux problems are MUCH easier to solve).

> Will I run into a maximum on the number of processes that can run?
> I have about 45-60 days to do this, Should I be waiting for the 2.2
> Kernel
> or do you think the latest 2.0.xx kernel can handle it?

Personally, I think you should hold out for 2.2 if possible. There are
just so many improvements..... But.. I do think that 2.0 could work for
you. There are some 2.0 advantages (which will prob not live long after
2.2 but) like there is a patch to allow processes to have upto 30k file
descriptors.

I think the 2.0 kernles used a hardware context switch that was limited to
4096 processes.. Though there are prob patches to get rid of this limit.

> I was hesitant about posting to the kernel list, but I figure that it is
> the
> kernel that has to handle the load, and I really need to know that Linux
> can
> handle this before I convince the customer to spend the money on the
> hardware.

:)

> If this works, I have several other customers who would also be good
> candidates for a Big Ass Linux Machine.
>
> Thanks for any suggestions you may have.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/