Why does 4KB Random Read suck @ Vertex 3 MAX IOPS?

[Axonn]

http://www.anandtech.com/show/4341/ocz-vertex-3-max-iops-patriot-wildfire-ssds-reviewed/4

See chart #3 on the page: Desktop Iometer - 4 KB Random Read (4K aligned).

I'm wondering why everybody praises SandForce so hard when it has that horrific RR performance.

Isn't the READ which is most important when loading programs?

I'm aware that it's RANDOM Read, but after all, SSDs will always become fully randomized since you can't run Defrag on them (it would be wasting cells). That is, if Defrag even works.

[not a SSD owner yet - got dual Raptors since 2007]

 

[EarthwormJim]

I guess the Sandforce controller has a harder time decompressing the data for reads vs. compressing it for writes.

A large portion of it's speed advantage over competing drives is due to the fact that it literally writes less data to the drive thanks to compression.


4k reads make up a small % of OS usage. 4k or 8k writes make up the largest % at least on my computer. You can use diskmon to determine what your usage pattern is like. Most programs read are not so tiny either.

 

[Seero]

http://www.anandtech.com/show/4341/ocz-vertex-3-max-iops-patriot-wildfire-ssds-reviewed/4

See chart #3 on the page: Desktop Iometer - 4 KB Random Read (4K aligned).

I'm wondering why everybody praises SandForce so hard when it has that horrific RR performance.

Isn't the READ which is most important when loading programs?

I'm aware that it's RANDOM Read, but after all, SSDs will always become fully randomized since you can't run Defrag on them (it would be wasting cells). That is, if Defrag even works.

[not a SSD owner yet - got dual Raptors since 2007]

My guess is this:

Say you have 10 dollars worth of quarters and we live 10 miles apart. Which of the following methods is faster.
Method 1:
I come to you and take 1 quarter, then go back home and come back again for another quarter until I eventually got 10 bucks.

Method 2:
I come to you and take 40 quarters, then go back home.

The time required to actually search your house is extremely short compare to the trip. In fact, it doesn't matter of the quarters are together or not as the time required for me to find them is exactly the same because I knew where those quarters are and capable of retrieving 5-10 quarters at once.

 

[PowerYoga]

I thought its because of the compression that sandforces are slow in read. A normal SSD typically has a much faster reads compared to any type of magnetic drive because there's no seek time for the spindle.

 

[Seero]

I thought its because of the compression that sandforces are slow in read. A normal SSD typically has a much faster reads compared to any type of magnetic drive because there's no seek time for the spindle.

If you compare it with other SSDs, the compression may have something to do with it, but 30+mb/sec @ 4k read is not slow. 10k rpm HDD can do 6.8mb/sec.

 

[Axonn]

EarthwormJim: I thought decompressing is faster than compressing.

PowerYoga: you're probably talking about access time?

Seero: Your "Hi grandma"-type example is confusing. You should hit me with the technical shovel, I'll understand better ::- D. The quarters should be... what? 4 KB = 4096 bytes.

Also, my question wasn't related to 4 KB read for ALL drives, but why is SandForce so slow with it?

Ok, I know it's still fast compared to a HDD, but I always thought decompressing works faster than compressing ::- D.

 

[Idontcare]

http://www.anandtech.com/show/4341/ocz-vertex-3-max-iops-patriot-wildfire-ssds-reviewed/4

See chart #3 on the page: Desktop Iometer - 4 KB Random Read (4K aligned).

I'm wondering why everybody praises SandForce so hard when it has that horrific RR performance.

Isn't the READ which is most important when loading programs?

I'm aware that it's RANDOM Read, but after all, SSDs will always become fully randomized since you can't run Defrag on them (it would be wasting cells). That is, if Defrag even works.

[not a SSD owner yet - got dual Raptors since 2007]

I've wondered this myself but have assumed it is nothing more than a tradeoff that is being made in the firmware parameters. RR@4K is slow because the RR or RW or SR or SW at some other interval is boosted all the more.

Why else would a standard vertex 3 240GB outperform its max IOPs counterpart? Has to be nothing more than just firmware optimization IMO.

 

[Seero]

EarthwormJim: I thought decompressing is faster than compressing.

PowerYoga: you're probably talking about access time?

Seero: Your "Hi grandma"-type example is confusing. You should hit me with the technical shovel, I'll understand better ::- D. The quarters should be... what? 4 KB = 4096 bytes.

Also, my question wasn't related to 4 KB read for ALL drives, but why is SandForce so slow with it?

Ok, I know it's still fast compared to a HDD, but I always thought decompressing works faster than compressing ::- D.

Technical it is then.
Sandforce controller write least data on flash than actual, some sees it as compression. Doesn't matter if it is compression or decompression, both take time. The idea isn't to save time, but the need to write on flash. Although that also means less actual read, the algorithm which maps it back to the original data does takes time. The read + map is what makes vertex 3 MAX slower than the high end SSD, not random read. That means, the dominate factor on SSD performance isn't on how data are distributed on the flash, but the amount of data retrieve/store + mapping.

Depends on the actual mapping algorithm, it may take the same amount of time to map 4k compare to 128k. It may not take the same amount of time, but 32x the time needed to map 4k data must be greater than to map 128k data, meaning 4k read will be slower than 128k read. That is, regardless of the distribution of how those 128k data is, the time require to get them with 1 pass is faster than 32 passes. I haven't factor in I/O controller and CPU delays on multi pass.

The funny part is this, with mapping, the actual access to flash is reduced, meaning the time on accessing flash is reduced. It is hard to say rather the map actual slows down the read/write process. In 128 read/write, vertex 3 MAX actually did as good or even better than others, meaning that the time saved from accessing flash > the extra time needed to map. Of course, since I don't have access to the actual code, so it is nothing but a claim.

 

[Axonn]

Idontcare: Thank you for attaching the picture ::- ). It is possible, indeed, that this is a firmware optimization. As far as we all see in the OTHER charts, Vertex 3 MAX IOPS works very well. And besides, that's only a bench, not REAL world performance.

Seero: that's more like it ::- D. Thank you. I got it now. The mapping thing makes sense. And besides, as EarthwormJim said, most of the stuff which goes on under the OS uses reads which are greater than 4 KB. Same with apps'n'games.