random write experiment with IO-meter

From OpenSSDWiki
Viewed 5311 times, With a total of 4 Posts
Jump to: navigation, search

Just Got Here
Threads 1
Posts 2
I conducted a random write experiment with IO-meter.
but I got a strange result.
below picture is the result.
I think that the maximum performance at 128Kbyte results from cahnnel and way effect.
but I don't understand performance drop at 512kbyte and 1024kbyte.
do you know why I got that result?

- I used both TutorialFTL and GreedyFTL. And I got same results.

Clicked A Few Times
Threads 11
Posts 39
I think its due to the default configuration of the jasmine board only having 8 banks (2CH-4WAY).
By default the Jasmine board uses a 32KB page size, 32KB * 8 Banks = 128KB.
Once all 8 banks are in use the FCP won't accept new commands until one of the banks is free.

You can play with different read/write techniques and see which yields the best performance though.
In the default configuration:
CH0 = Banks 0,2,4,6
CH1 = Banks 1,3,5,7
Edited On 8:49:31 PM - Mon, Jun 13th 2011 by Jeremy Brock
Jeremy Brock

A+ Perfect Computers (http://www.recovermyflashdrive.com)
President / Sr. R&D Engineer
2465 Black Rock Tpke., STE. C
Fairfield, CT 06825, USA
Tel: 203-549-9761
Email: jeremy@recovermyflashdrive.com

Just Got Here
Threads 1
Posts 2
but I think that the NAND flashs can be fully utilized at 512kbyte and 1024kbyte as at 128Kbyte,
becuase the FCP can accept new commands and issue that command to NAND as soon as the banks is free.
so I think the performance at 512kbyte and 1024kbyte should be simular to performance at 128Kbyte.
don't you think so?
Edited On 2:34:14 AM - Sat, Jun 18th 2011 by Jpark33

Clicked A Few Times
Threads 11
Posts 39
FCP is blocked by both the channel and bank, the I/O bus is shared by all banks on the channel.

Only one READ/WRITE transmission can occur per bank per channel, the FCP's trick is not to block the bus while NAND is busy "thinking". During this time it will issue commands on another bank until NAND is ready to send data or accept new commands. This increases performance by ~25% because time on the bus isn't wasted.

Barefoot Fig5.png

eg: NAND Speed * 1.25 x number of channels = throughput

Please correct me if I'm wrong.
Also as I understand RB is shared between Banks 00/16 & 04/20, I don't understand how the controller knows which bank is ready, but it works.. so I don't ask questions.
Edited On 9:20:05 AM - Sat, Jun 18th 2011 by Jeremy Brock

Clicked A Few Times
Threads 0
Posts 17
Jpark33, I have two questions:

1. What was the boundary of alignment? In iometer you can set "align I/Os on..."

2. Did you see similar degradation with sequential writes, too?

Forum >> Jasmine OpenSSD Platform >> Jasmine General Issues

Who's here now Members 0 Guests 0 Bots/Crawler 0

AWC's: 2.5.12 MediaWiki - Stand Alone Forum Extension
Forum theme style by: AWC
Personal tools