On Tue, Jan 31, 2006 at 10:01:33AM -0500, stan wrote:
> I'm going to upgrade our Amanda tape/index server.
>
> I've narrowed the choice of machines down to one of 2 Sun Models. Either a
> W2100Z, or a Ultra 40. I plan on using an Ultrim tape drive, and doing
> compression on the server, so I need lot's of CPU power, and of course high
> I/O throughput.
>
> I was wondering what people thought about this selection?
>
> The W2100Z is a SCSI machine, but I would still add a 2nd SCSI card
> dedicated to the tape drive. The Ultra 40 looks like it would have more CPU
> power, as the processors are dual core. I plan on using Solaris 10, if that
> matters. Is SATA equal to SCSI for the disk I/O?
My first impression was SATA is fine for holding and 4 threads
of the Ultra 40 was a big plus. But I decided to ask someone,
Al Hopper, from the Solaris x86 mailing list whose opinion I value.
Al concurred that the Ultra 40 is the better choice and brought
up a number of points in support of his decision that I'd not
considered. Things like end of product life and benefits of
moving to Sun's new filesystem, ZFS, in the near future.
Its a little off topic, but I thought my question and Al's reply
worth posting.
On Tuesday, Jan 31 2006, Al Hooper replied:
On Tue, 31 Jan 2006, Jon LaBadie wrote:
>
> > Al,
> >
> > On another mailing list, dealing with "amanda" backups
> > a poster asked about choices for a backup server. He
> > was questioning the Ultra 40 vs the W2100z. Both would
> > be dual processors, the Ultra 40 dual core as well, but
> > at a slightly lower clock speed.
> >
> > Assuming you are unfamiliar with amanda, the backup server
> > would be receiving multiple, simultaneous network connections
> > receiving tar data. Each connection would be piped through
> > two processes (one a gzip) and collected on a "holding disk".
> > This holding disk would be SCSI on the W2100z, SATA on the
> > Ultra 40.
> >
> > Dumps that have completly collected on the holding disk are
> > queue for transfer to tape, one at a time. This would be
> > occuring at the same time as the collection of other dumps.
> > This involves two other processes, but is essentially a "dd"
> > to tape. The tape drive will be an LTO (2 or 3?) connected
> > to its own add-in SCSI controller. LTO drives today seem to
> > be able to saturate a controller's I/O capability.
> >
> > My intuition says the SATA drives of the Ultra 40 should be
> > fine for collecting the data, so lack of SCSI here is no problem.
> > And the ability to run 4 slightly slower threads vs 2 faster
> > ones in the W2100z could be a big factor for lots of gzip'ping.
> >
> > Do you see anything strongly pushing one solution over another?
>
> Yes! But you knew that already! :)
>
> ZFS will ship with the next Solaris 10 Update - that'll be
> Update 2, and it should ship "on-time". It's just possible
> (looking in my crystal ball) that you'll be able to boot off
> ZFS in update 2 - because I've seen blogs that are 2 months+
> old saying they had ZFS booting. At that point in time - using
> anything other than cheap fast SATA disk drives just won't make
> sense IMHO. SCSI is EOL technology - as I've pointed out
> before on the list - the primary reason is that commands are
> sent over the bus in s-l-o-w legacy xfer rate (5Mb/Sec) 8-bits
> wide[0]. And since it takes on average 5 SCSI commands to do
> anything useful, you simply *cannot* get enough commands to the
> disk drive (even a single disk drive on a dedicated SCSI bus)
> to keep the disk drive busy. Therefore, you're wasting your
> money on SCSI and buying EOL technology.
>
> ZFS, even in its current, untuned, and CPU intensive form, will
> give you better performance - even though it won't saturate the
> SATA drives at this time. The next rev, after the code has
> been optimized, will probably double throughput IMHO (WAG
> etc). But in the first rev, right now via Solaris Express,
> it'll easily do 100M bytes/Sec[1] without even trying.
>
> In the meantime, assuming you won't be able to boot off ZFS,
> just boot off one SATA (dedicated OS) drive and you'll have 3
> other SATA drives to play with using regular SVM - or ZFS if
> you want to load the next Solaris express release on the new
> box (build 31 or 32 should ship in the next week). When ZFS
> ships, use all 4 SATA drives in a raidz config for everything
> (data, OS, swap, home etc). This may sound conter-intutive -
> using a 4-drive raidz for everything, including your backup
> processes, but that is exactly what ZFS is designed to do.
> Also don't forget that you can setup a ZFS filesystem with
> compression enabled - so you could spool your incoming data
> stream to a ZFS compressed filesystem - then your read
> performance would be 2x - depending on the compressibility of
> your incoming data stream.
>
> Since the W1100 and W2100 are based on designs made by someone
> else, they'll be EOLed ASAP AFAICT. This is my personal
> opinion only. The Ultra40 is Suns long term solution to a
> deskside dualie tower and it will be better supported going
> forward IMHO. You should be selling W[12]00Zs right now! :)
>
> PS: I know Sun is working on drivers for some dumb 4 & 8
> channel SATA controllers made by someone else. The SuperMicro
> board is one of the target boards. When these drivers ship,
> you'll be able to connect to a dumb 4-slot external disk drive
> cage and now you'll have another ~ 2Tb (4 * 500Gb) of disk
> drives to play with. And another raidz device on separate
> physical spindles and the ZFS tools to slice/dice them into
> many filesystems.
>
> PS: Think about where ZFS is headed. Assuming a future tuned
> rev, say 6 months from now (or less), you could take an input
> stream of ~ 100Mb/Sec from a backup client, store it on a
> compressed raidz filesystem, then read it at 200Mb/Sec from a
> 4-way raidz device - but since its already compressed, so you
> can expect to see 2 * 200Mb/Sec reads - which is now ready for
> compression via gzip, or whatever the compression method you
> use with Amanda, and it'll easily saturate your backup device
> and keep it streaming.
>
> PPS: Don't forget that WD just shipped their 150GB 10,000 RPM
> SATA Raptor with Native Command Queuing (NCQ). This drive will
> seriously kick SCSIs ask, in terms of throughput and I/O
> Ops/Sec. It is available in small quantities right now.
>
> I'm assuming you have not played with ZFS yet? It's innovative,
> disruptive and compelling technology. After you play with ZFS
> you won't want to go "backwards".
>
> [0] Which is why we have the move to Serial Attached SCSI. Not
> because they wanted to disrupt a perfectly good, stable,
> revenue generating SCSI market!
>
> [1] Did you see Bart Smalders recent post? Getting 100Mb/Sec
> with a low-end AMD 3200+ (IIRC) and some cheap SATA drives.
>
> Regards,
>
> Al Hopper Logical Approach Inc, Plano, TX. al AT logical-approach DOT com
> OpenSolaris.Org Community Advisory Board (CAB) Member - Apr 2005
--
Jon H. LaBadie jon AT jgcomp DOT com
JG Computing
4455 Province Line Road (609) 252-0159
Princeton, NJ 08540-4322 (609) 683-7220 (fax)
|
