New SSD technology, new problems

[IntelUser2000]

As you guys might know, first generation SSDs have suffered poor overall write problems, and the following generations revealed poor random write problems. They also suffered reliability problems. SSD based systems randomly failed much before their MTBF numbers.

Now SSD market is starting to mature and new SSDs are coming out. New SSDs which promise astonishing transfer rates(ex. FusionIO) and amazing random writes(ex. Intel SSD). So problems like "stuttering" are gone right and everything is alright in the SSD world??

New technology is bringing new problems....

FusionIO, a super fast PCI-Express based SLC SSD card featuring 725MB/sec read and 400MB/sec writes, with very good random write 4k figures in Crystaldiskmark.

http://forum.ssdworld.ch/viewtopic.php?f=4&t=58

run a heavy server load and...

http://forum.ssdworld.ch/viewtopic.php?f=1&t=59

Performance drops drastically.

X25-E, Intel's SLC based SSD drive featuring 250MB/sec read and 170MB/sec write(270MB read/200MB write Crystaldiskmark).

http://forum.ssdworld.ch/viewtopic.php?f=4&t=81

run a heavy server load like IOMeter and...

http://forum.ssdworld.ch/viewtopic.php?f=1&t=82

Performance also drops substantially.

Possible reasons:

"Here's what's going on. The cards get slower the more you write to them. In the background a "grooming" process is doing some kind of cleanup (obviously erasing empty blocks, but perhaps something else as well)."

My theory(advancing on what the poster said above):

These SSDs have very advanced write-levelling algorithms and very low write amplification to offer the BEST reliability possible. As both X25-E and FusionIO are targeted for servers, it makes sense. The "grooming" process the controller is doing is making sure that every block is written evenly so no random failures occur.

The Intel MLC shows a similar problem because the controller used is exactly same as the SLC version. The Intel controller is supposed to have an advanced implementation and algorithm that it will regain the performance back.

My test with X25-M MLC Intel:

"1st one, Oct 23, 2008
http://img101.imageshack.us/my...age=hdtunex25m2no1.jpg

2nd one November 4, 2008:
http://img293.imageshack.us/my...dtunex25m3laterki5.jpg

Test I did December 10, 2008:
http://img525.imageshack.us/my.php?image=x25mcdmrp0.jpg

As you can see on the last pic, the dip shown at 20% disappeared after prolonged usage.

Crystaldiskmark
http://img525.imageshack.us/my.php?image=x25mcdmrp0.jpg"

So from both the review and my tests the recovery process does work. Now the question is how fast does the controller gain the performance back?

 

[IntelUser2000]

PC users shouldn't experience such a big slowdowns so often as the usage isn't as intensive. IOMeter stresses the drive a lot. Server users might want to look up on it more.

(To Intel's credit, the IOPS figure of the X25-E isn't too much far away from FusionIO despite less than half the sequential read/write figures. The degradation is also less. Yes, the controller is pretty good)