[BBLISA] Dell/other Linux-raid-NFS-server, ready for primetime?

[Eddy Harvey]

I have a fair amount of experience supplying data to engineers on their
engineering systems.  Sometimes NAS, sometimes DAS.  Sometimes SAN.

This is what I found - About 3 years ago, I bought a linux machine with 1TB
storage, plugged it into the GB ether.  Started sharing NFS.  I found that
NFS itself adds about 50-90% overhead.  Severely killing throughput compared
to DAS or SAN.  I found that the Linux kernel (out of the box) handled NFS
well as long as it was reasonable demand.  But the engineers of course use
every system at maximum throttle.  The kernel would frequently (once a week
or so) suffer some kind of buffer overflow, and halt or reboot.  (yikes!)
So the linux kernel needed some tweaking.

I got a company (forget their name now) that layered on extra software to
make it more stable.  They were able to add on an $8k software product that
made the system faster and more stable.  But the amount of work to set up,
the total end cost, and blah blah ... Made the whole system a bad decision
in retrospect.

I had a netapp.  Obviously worked like a champ.  Fast, reliable, swiss army
knife.  But the pricetag, oh the pricetag.

All of this is still limited to 1Gb due to the ether.  (Well, of course,
unless you have a 10Gb switch.)

At present, I do this - I have the core apps, tools, sitting on an NFS
share, on Gb ether.  I give each machine a non-backed up /scratch area,
which is DAS.  This seems ideal.  You get the 3Gb performance and
hassle-less, low-overhead of DAS, which is cheap.  You get the core tools
backed up.  You tell the engineers to store their source code on the NFS
share (to be backed up) and store their simulation results on /scratch.

So far so good. I think it's a winning combination.




-----Original Message-----
From: bblisa-bounces at bblisa.org [mailto:bblisa-bounces at bblisa.org] On Behalf
Of Doug Mildram
Sent: Thursday, November 09, 2006 12:06 PM
To: bblisa at bblisa.org
Subject: [BBLISA] Dell/other Linux-raid-NFS-server, ready for primetime?

doug>   sigh, it has been years since I tried-and-failed....
putting a linux-based NFS server into a high-traffic engineering farm.
(pesky HW engineers and their simulation or design data).
I ended up putting freeBSD on the $10k IDE-raid box I bought in 2001.
Can't do the freeBSD thing for a "new requirement" on a LOWER!? budget, so:

THE QUESTION: Now+again, I wonder who has tried and failed/succeeded with
cheap (linux) NFS server... ?

ok, to do what?
oh, just read a few and write a lot of GB in parallel (multi-NFS-clients).
e.g. lots of files (100k here, 100k there per hour...stay the course...) 

Is any linux distro better than another for this? 
RHEL ? (or whiteboxEL which works the same, right?) other?

Don't need to reminded of obvious BETTER solutions ...
I'm trolling for "what works great new-and-under-$5k".

The recommendation I'm trying to fight for= NetApp. Cmon, $20k is nothing!)

Sorry this is long.  thanks for listening/replies.  Gory detail below.

How did it fail me on (redhat-before-FC; tried 2.5 kernel, no better...)?
Despite giga-memory and excess I/O avail for enet+disk, what puked was NFS
service. I bumped up nfsd #processes, tried various wsize,rsize, and sure
it'd WORK, but add a few more busy NFS clients, and the server seemed to
self-destruct. Symptoms: 20+ load, crawling NFS response.

Sorry, my "requirements" and "symptoms" are imprecise, but it only seemed to
take a few hours and a few McWhoppers, i mean a couple gb mostly-written by
a couple users IN a few hours, to CHOKE/gag/pretty-much-hang Linux.  Havent
bought RHEL...too wary I am?
RH/RedHat (nfs client!) systems do "play nicely" otherwise here, though.
Network thruput can be 20gb-30gb/hour on 100mb enet, IF you don't use NFS.

What DOES work (NFS server) without choking, in my budget+history, has been:
      --NetApp  (which I still recommend. Boss says $18k==toomuch).
                    (I say: a steal! Price has halved every 3? yrs!)
      --freeBSD  (not the perfect fit either for a remote-site-main-"filer")

Here's a possible HW platform, if the OS loaded onto it was appropriate:
Dell PE1900 poweredge w/optional integrated SAS/SATA raid5,
       5x500gb SATA 7.2krpm, probably close to 2tb usable,
       Broadcom NetXtreme gb-nic (something I havent GOTTEN/had yet),
       (2) 2.0ghz dual core xeon 5130 4mb-cache, 1333mhz FSB,
   $2900+tax+ship=$3200 or $3300.   No OS.    (OS is my question/problem,
       or you could say lack of budget for an $18k 1+TB NetApp is my
problem.)

Perhaps this troll will perk up my energy/luck.  
If given a good tip, I WILL try free ideas/OS/tweaks first :) Impossibly, I
must figure out if it will work WELL before buying anything.


_______________________________________________
bblisa mailing list
bblisa at bblisa.org
http://www.bblisa.org/mailman/listinfo/bblisa