The Performance Degradation Problem

When Intel first released the X25-M, Allyn Malventano discovered a nasty corner case where the drive would no longer be able to run at its full potential. You basically had to hammer on the drive with tons of random writes for at least 20 minutes, but eventually the drive would be stuck at a point of no return. Performance would remain low until you secure erased the drive.

Although it shouldn't appear in real world use, the worry was that over time a similar set of conditions could align resulting in the X25-M performing slower than it should. Intel, having had much experience with similar types of problems (e.g. FDIV, Pentium III 1.13GHz), immediately began working on a fix and released the fix a couple of months after launch. The fix was nondestructive although you saw much better performance if you secure erased your drive first.

SandForce has a similar problem and I have you all and bit-tech to thank for pointing it out. In bit-tech's SandForce SSD reviews they test TRIM functionality by filling a drive with actual data (from a 500GB source including a Windows install, pictures, movies, documents, etc...). The drive is then TRIMed, and performance is measured.

If you look at bit-tech's charts you'll notice that after going through this process, the SandForce drives no longer recover their performance after TRIM. They are stuck in a lower performance state making the drives much slower when writing incompressible data.

You can actually duplicate the bit-tech results without going through all of that trouble. All you need to do is write incompressible data to all pages of a SandForce drive (user accessible LBAs + spare area), TRIM the drive and then measure performance. You'll get virtually the same results as bit-tech:

AS-SSD Incompressible Write Speed
  Clean Performance Dirty (All Blocks + Spare Area Filled) After TRIM
SandForce SF-1200 (120GB) 131.7MB/s 70.3MB/s 71MB/s

The question is why.

I spoke with SandForce about the issue late last year. To understand the cause we need to remember how SSDs work. When you go to write to an SSD, the controller must first determine where to write. When a drive is completely empty, this decision is pretty easy to make. When a drive is not completely full to the end user but all NAND pages are occupied (e.g. in a very well used state), the controller must first supply a clean/empty block for you to write to.

When you fill a SF drive with incompressible data, you're filling all user addressable LBAs as well as all of the drive's spare area. When the SF controller gets a request to overwrite one of these LBAs the drive has to first clean a block and then write to it. It's the block recycling path that causes the aforementioned problem.

In the SF-1200 SandForce can only clean/recycle blocks at a rate of around 80MB/s. Typically this isn't an issue because you won't be in a situation where you're writing to a completely full drive (all user LBAs + spare area occupied with incompressible data). However if you do create an environment where all blocks have data in them (which can happen over time) and then attempt to write incompressible data, the SF-1200 will be limited by its block recycling path.

So why doesn't TRIMing the entire drive restore performance?

Remember what TRIM does. The TRIM command simply tells the controller what LBAs are no longer needed by the OS. It doesn't physically remove data from the SSD, it just tells the controller that it can remove the aforementioned data at its own convenience and in accordance with its own algorithms.

The best drives clean dirty blocks as late as possible without impacting performance. Aggressive garbage collection only increases write amplification and wear on the NAND, which we've already established SandForce doesn't really do. Pair a conservative garbage collection/block recycling algorithm with you attempting to write an already full drive with tons of incompressible data and you'll back yourself into a corner where the SF-1200 continues to be bottlenecked by the block recycling path. The only way to restore performance at this point is to secure erase the drive.

This is a real world performance issue on SF-1200 drives. Over time you'll find that when you go to copy a highly compressed file (e.g. H264 video) that your performance will drop to around 80MB/s. However, the rest of your performance will remain as high as always. This issue only impacts data that can't be further compressed/deduped by the SF controller. While SandForce has attempted to alleviate it in the SF-1200, I haven't seen any real improvements with the latest firmware updates. If you're using your SSD primarily to copy and store highly compressed files, you'll want to consider another drive.

Luckily for SandForce, the SF-2500 controller alleviates the problem. Here I'm running the same test as above. Filling all blocks of the Vertex 3 Pro with incompressible data and then measuring sequential write speed. There's a performance drop, but it's no where near as significant as what we saw with the SF-1200:

AS-SSD Incompressible Write Speed
  Clean Performance Dirty (All Blocks + Spare Area Filled) After TRIM
SandForce SF-1200 (120GB) 131.7 MB/s 70.3 MB/s 71 MB/s
SandForce SF-2500 (200GB) 229.5 MB/s 230.0 MB/s 198.2 MB/s

It looks like SandForce has increased the speed of its block recycling engine among other things, resulting in a much more respectable worst case scenario of ~200MB/s.

Verifying the Fix

I was concerned that perhaps SandForce simply optimized for the manner in which AS-SSD and Iometer write incompressible data. In order to verify the results I took a 6.6GB 720p H.264 movie and copied it from an Intel X25-M G2 SSD to one of two SF drives. The first was a SF-1200 based Corsair Force F120, and the second was an OCZ Vertex 3 Pro (SF-2500).

I measured both clean performance as well as performance after I'd filled all blocks on the drive. The results are below:

6.6GB 720p H.264 File Copy (X25-M G2 Source to Destination)
  Clean Performance Dirty (All Blocks + Spare Area Filled) After TRIM
SandForce SF-1200 (120GB) 138.6 MB/s 78.5 MB/s 81.7 MB/s
SandForce SF-2500 (200GB) 157.5 MB/s 158.2 MB/s 157.8 MB/s

As expected the SF-1200 drive drops from 138MB/s down to 81MB/s. The drive is bottlenecked by its block recycling path and performance never goes up beyond 81MB/s.

The SF-2000 however doesn't drop in performance. Brand new performance is at 157MB/s and post-torture it's still at 157MB/s. What's interesting however is that the incompressible file copy performance here is lower than what Iometer and AS-SSD would have you believe. Iometer warns that even its fully random data pattern can be defeated by drives with good data deduplication algorithms. Unless there's another bottleneck at work here, it looks like the SF-2000 is still reducing the data that Iometer is writing to the drive. The AS-SSD comparison actually makes a bit more sense since AS-SSD runs at a queue depth of 32 and this simple file copy is mostly at a queue depth of 1. Higher queue depths will make better use of parallel NAND channels and result in better performance.

Sequential Read/Write Speed AnandTech Storage Bench 2011: Much Heavier
POST A COMMENT

144 Comments

View All Comments

  • jwilliams4200 - Friday, February 18, 2011 - link

    In that case, it would be helpful to print two after-TRIM benchmarks: (1) immediately after TRIM and (2) steady-state after-TRIM (i.e., TRIM, let the drive sit idle for long enough for GC to complete, then benchmark again) Reply
  • jcompagner - Thursday, February 17, 2011 - link

    what i never understood or maybe i should read a bit more the previous articles, is that how come that a SSD can write many times faster then it can read?
    It seems to me that read is way easier to do then write...
    Reply
  • vol7ron - Friday, February 18, 2011 - link

    I originally thought that, but SSDs first write to the controller, which organizes the data for storing it to the disk. The major point is that the data can go anywhere in the array of NAND nodes and the list of the next available node in the stack is available almost immediately, whereas a read requires a hash lookup of where the data is stored, which means the seek could take longer to accomplish.

    I, as well, am not certain that's true, but that's my best guess.
    Reply
  • AnnihilatorX - Saturday, February 19, 2011 - link

    Only for Sandforce controllers.
    Sandforce compresses the incoming data at real time. If the incoming data is highly compressible, in a very extreme example, writting a 500MB blank text file, will be instantaneous. So you see 500MB/ms or something ridiculous.

    It is also possible for write speeds to exceed read in burst when small amount of data is written to DRAM on other controllers
    Reply
  • Soul_Master - Thursday, February 17, 2011 - link

    For zero impact from source performance, I suggest to copy data from RAM drive to your test hard disk. Reply
  • Anand Lal Shimpi - Thursday, February 17, 2011 - link

    That's a great suggestion. I ran out of time before I left the country but I'll be playing with it some more upon my return :)

    Take care,
    Anand
    Reply
  • MrBrownSound - Thursday, February 17, 2011 - link

    I think the intel x25m was a pretty good control group to send the data from. I would auctally like to see the changes when sending the data through the RAM; that would be interesting. Reply
  • Hacp - Thursday, February 17, 2011 - link

    Anand,
    You still direct your readers to your Vertex2 article but OCZ has changed its performance on those drives. Your results are no longer valid and it would be dishonest to link the old Vertex2 performance numbers in this new article when they do not reflect the new slower performance of the Vertex2 today.
    Reply
  • Anand Lal Shimpi - Thursday, February 17, 2011 - link

    I've seen the discussion and based on what I've seen it sounds like very poor decision making on OCZ's behalf. Unfortunately my 25nm drive didn't arrive before I left for MWC. I hope to have it by the time I get back next week and I'll run through the gamut of tests, updating as necessary. I also plan on speaking with OCZ about this. Let me get back to the office and I'll begin working on it :)

    As far as old Vertex 2 numbers go, I didn't actually use a Vertex 2 here (I don't believe any older numbers snuck in here). The Corsair Force F120 is the SF-1200 representative of choice in this test.

    Take care,
    Anand
    Reply
  • Quindor - Thursday, February 17, 2011 - link

    Good to hear that you are addressing the problems surrounding the Vertex 2 drives. There aren't many websites out there which deliver well thought through reviews and bechmarks such as Anandtech does, although some are getting better.

    I did some benchmarks on my own and with the new 25nm NAND the new 180GB OCZ Vertex2 can actually be slower then my more then a year old 120GB OCZ Vertex1.

    If anyone is interested. They can find an overview of the benchmarks performed on the following page. https://picasaweb.google.com/quindor/Benchmarks#

    Still, I would love to see an in depth comparsion as you are famous for. ;)

    For my personal usage scenario (my own ESXi server), the speed decrease will be of minimal effect because running multiple template cloned guests, the dedup and compression should be able to do their work just fine. ;)
    Reply

Log in

Don't have an account? Sign up now