[Tig] St. Diig

[timothy norman huber]

On Oct 18, 2007, at 3:50 PM, Bob Friesenhahn wrote:
> Regardless of rapidly-growing CPU performance, I/O is the true
> throughput bottleneck and the vendor who most effectively distributes
> and balances the I/O (while preserving convenience) is the one who
> will deliver the most performance per seat.


We're working on the I/O bottleneck at Violin.

Right now we've got 1/2 TB of DRAM in 2U connecting via PCIe.   
Random  I/O performance is the same as Linear I/O because DRAM isn't  
seek-bound like HDD technology.   Soon we will support 5TB of Flash  
in the same appliance, and you will be able to intermix Flash and  
DRAM in the same box to tweak the performance characteristics to your  
liking.

1/2 TB in 2U is very dense, and the only way we can get that much in  
the box without melting cables is by solving the heat problem.  By  
solving that, we also reduced power consumption 90% over server memory.

Random I/O kills even the fastest HDD technology.  15k drives have  
been around a long time, yet they're only gaining in capacity, not in  
performance.  Seek times are 5ms at best, which means that for random  
access, they are seek-limited.

Here's an example.

Access patterns-performance disk:

bandwidth = ~100MB/s

IOPS = I/O Per Second

Linear access for 8k block size example
IOPS = Bandwidth/ block size  = 100MB/ [8k block size] = 12.5k IOPS

Random access are seek-limited
seek time = ~5ms

IOPS limited by seek time or ~ 200 IOPS
Bandwidth =  [8k block size] * 200 IOPS = 1.6MB/s


Access patterns -performance disk arrays

10 disk array-Linear accesses will scale

Linear bandwidth 	= 1GB/s
Linear IOPS 			=  1GB/[8k block size]	= 125k

Random accesses over data sets exceeding disk, OS and array cache are  
seek-limited

seek time 			= 5ms
IOPS 				= 200 IOPS
Random access bandwidth = 200 * [block size]
Random access for 8k block size example = 1.6MB

Now imagine if you're using 1k block sizes with random I/O.  Its even  
more pathetic.

We're shipping our first production units hopefully next week , and  
just recently exited stealth mode at LinuxWorld in August.

T




Timothy Norman Huber
Violin Memory, Inc
33 Wood Ave South
Iselin, NJ, 08830


http://www.violin-memory.com/index.cfm
timothyhuber at violin-memory.com


(Los Angeles)
640 Woodlawn Ave
Venice, CA 90291

310 795-6599 Cell