Skip site navigation (1)Skip section navigation (2)
Date:      Mon, 01 May 2000 17:35:57 -0700
From:      David Greenman <dg@root.com>
To:        Mike Smith <msmith@FreeBSD.ORG>
Cc:        freebsd-stable@FreeBSD.ORG
Subject:   Re: How good is AMI MegaRAID support? 
Message-ID:  <200005020035.RAA28815@implode.root.com>
In-Reply-To: Your message of "Mon, 01 May 2000 17:20:57 PDT." <200005020020.RAA04531@mass.cdrom.com> 

next in thread | previous in thread | raw e-mail | index | archive | help
>>    TeraSolutions' RAID systems (TSR-2200) as used on ftp.freesoftware.com are
>> capable of >9,000 IOPS.
>
>Is this a controller capability, or does it represent a sustainable load 
>over the entire array?  How do you measure this? (I'd love to add to my 
>benchmark/test suite).

   The controller itself is spec'd at 10,000 IOPS out of the cache, but under
FreeBSD a more practical limit is about 8,500 due to OS latency and SCSI bus
overhead issues.
   As for how it's measured, usually you write the same block over and over
again (to test write-to-cache performance), or read the same block over and
over again. You do this with 50 or so processes simultaneously to take
advantage of overlapped tagged operations.
   For read from the disk drives, the performance is pretty much whatever
the drives are rated at. We use the fastest 10K RPM drives on the market,
but the actual number you get depends on both the speed and number of the
drives in the array. Write performance is hard to measure since the cache
defers the writes.
   On a software RAID-5, write performance can be expected to be totally
lousy due to the lack of non-volatile write-back cache. If it's not lousy,
then your filesystem is in danger of being destroyed on power fail or system
crash. I don't think anyone who goes to the trouble of doing software RAID
is willing to risk losing everything due to a power failure.

>FWIW, most of the low-end PCI:SCSI RAID controllers claim throughput in 
>the 3-5k IOPs, and 20k is not an uncommon claim for mid-high end 
>controllers.  Simon Shapiro was pushing over 20k on the DPT Century 
>adapters in "real" applications.  I've had a hard time generating more 
>than 3k or so out of a FreeBSD box's I/O subsystem - we cluster so 
>aggressively that I typically run out of I/O bandwidth before I hit an 
>IOP limit.

   You need to use the raw (character) device for testing things like this.
RAID controllers certainly vary greatly in performance. I've tested just about
all of the SCSI-SCSI controllers on the market and can tell you that most
of them really suck. The Mylex DAC960-SX, for example, tops out at about
1500 IOPS; the Infortrend 3102U2G at about 3200 IOPS; the CMD CRD-5440 at
about 4400 IOPS; etc...
   Anyway, sorry, I didn't really want to get sucked into this discussion,
so I'll just jump out now as fast as I jumped in. :-)

-DG

David Greenman
Co-founder/Principal Architect, The FreeBSD Project - http://www.freebsd.org
Creator of high-performance Internet servers - http://www.terasolutions.com
Pave the road of life with opportunities.


To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-stable" in the body of the message




Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200005020035.RAA28815>