The Cleaning Lady and Write Amplification

Imagine you’re running a cafeteria. This is the real world and your cafeteria has a finite number of plates, say 200 for the entire cafeteria. Your cafeteria is open for dinner and over the course of the night you may serve a total of 1000 people. The number of guests outnumbers the total number of plates 5-to-1, thankfully they don’t all eat at once.

You’ve got a dishwasher who cleans the dirty dishes as the tables are bussed and then puts them in a pile of clean dishes for the servers to use as new diners arrive.

Pretty basic, right? That’s how an SSD works.

Remember the rules: you can read from and write to pages, but you must erase entire blocks at a time. If a block is full of invalid pages (files that have been overwritten at the file system level for example), it must be erased before it can be written to.

All SSDs have a dishwasher of sorts, except instead of cleaning dishes, its job is to clean NAND blocks and prep them for use. The cleaning algorithms don’t really kick in when the drive is new, but put a few days, weeks or months of use on the drive and cleaning will become a regular part of its routine.

Remember this picture?

It (roughly) describes what happens when you go to write a page of data to a block that’s full of both valid and invalid pages.

In actuality the write happens more like this. A new block is allocated, valid data is copied to the new block (including the data you wish to write), the old block is sent for cleaning and emerges completely wiped. The old block is added to the pool of empty blocks. As the controller needs them, blocks are pulled from this pool, used, and the old blocks are recycled in here.

IBM's Zurich Research Laboratory actually made a wonderful diagram of how this works, but it's a bit more complicated than I need it to be for my example here today so I've remade the diagram and simplified it a bit:

The diagram explains what I just outlined above. A write request comes in, a new block is allocated and used then added to the list of used blocks. The blocks with the least amount of valid data (or the most invalid data) are scheduled for garbage collection, cleaned and added to the free block pool.

We can actually see this in action if we look at write latencies:

Average write latencies for writing to an SSD, even with random data, are extremely low. But take a look at the max latencies:

While average latencies are very low, the max latencies are around 350x higher. They are still low compared to a mechanical hard disk, but what's going on to make the max latency so high? All of the cleaning and reorganization I've been talking about. It rarely makes a noticeable impact on performance (hence the ultra low average latencies), but this is an example of happening.

And this is where write amplification comes in.

In the diagram above we see another angle on what happens when a write comes in. A free block is used (when available) for the incoming write. That's not the only write that happens however, eventually you have to perform some garbage collection so you don't run out of free blocks. The block with the most invalid data is selected for cleaning; its data is copied to another block, after which the previous block is erased and added to the free block pool. In the diagram above you'll see the size of our write request on the left, but on the very right you'll see how much data was actually written when you take into account garbage collection. This inequality is called write amplification.


Intel claims very low write amplification on its drives, although over the lifespan of your drive a < 1.1 factor seems highly unlikely

The write amplification factor is the amount of data the SSD controller has to write in relation to the amount of data that the host controller wants to write. A write amplification factor of 1 is perfect, it means you wanted to write 1MB and the SSD’s controller wrote 1MB. A write amplification factor greater than 1 isn't desirable, but an unfortunate fact of life. The higher your write amplification, the quicker your drive will die and the lower its performance will be. Write amplification, bad.

Live Long and Prosper: The Logical Page Why SSDs Care About What You Write: Fragmentation & Write Combining
POST A COMMENT

296 Comments

View All Comments

  • GourdFreeMan - Tuesday, September 01, 2009 - link

    Yes, rewriting a cell will refill the floating gate with trapped electrons to the proper voltage level unless the gate has begun to wear out, so backing up your data, secure erasing your drive and copying the data back will preserve the life (within reason) of even drives that use minimalistic wear leveling to safeguard data. Charge retention is only a problem for users if they intend to use the drive for archival storage, or operate the drive at highly elevated temperatures.

    It is a bigger problem for flash engineers, however, and one of the reasons why MLC cannot be moved easily to more bits per cell without design changes. To store n-bits in a single cell you need 2^n separate energy levels to represent them, and thus each bit is only has approximately 1/(2^(n-1)) the amount of energy difference between states when compared to SLC using similar designs and materials.
    Reply
  • Zheos - Tuesday, September 01, 2009 - link

    Man you seem to know a lot about what you're talking about :)

    Yeah now i understand why SSD for database and file storage server would be quite a bad idea.

    But for personal windows & everyday application storage, seems like a pure win to me if you can afford one :)

    I was only worried about its life-span but thankx to you and you're quick replys (and for the maths and technical stuff about how it realy work ;) im sold on the fact that i will buy one soon.

    The G2 from Intel seems like the best choice for now but I'll just wait and see how it's going when TRIM will become almost enable on every SSD and i'll make my decision there in a couple of months =)


    Reply
  • GourdFreeMan - Wednesday, September 02, 2009 - link

    It isn't so much that SSDs make a bad storage server, but rather that you can't neglect to make periodic backups, as with any type of storage, if your data has great monetary or sentimental value. In addition to backups, RAID (1-6) is also an option if cost is no object and you want to use SSDs for long term storage in a running server. Database servers are a little more complicated, but SSDs can be an intelligent choice there as well if your usage patterns aren't continuous heavy small (i.e. <= 4K) writes.

    I plan on getting a G2 myself for my laptop after Intel updates the firmware to support TRIM and Anand reviews the effects in Windows 7, and I have already been using an Indilinx-based SLC drive in my home server.

    If you do anything that stresses your hard drive(s), or just like snappy boot times and application load times you will probably be impressed by the speeds of a new SSD. The cost per GB and lack of long term reliability studies are really the only things holding them back from taking the storage market by storm now.
    Reply
  • ninevoltz - Thursday, September 17, 2009 - link

    GourdFreeMan could you please continue your explanation? I would like to learn more. You have really dived deeply into the physical properties of these drives. Reply
  • GourdFreeMan - Tuesday, September 01, 2009 - link

    Minor correction to the second paragraph in my post above -- "each bit is only has" should read "each representation only has" in the last sentence. Reply
  • philosofool - Monday, August 31, 2009 - link

    Nice job. This has been a great series.

    I'm getting a SSD once I can get one at $1/GB. I want a system/program files drive of at least 80GB and then a conventional HDD (a tenth of the cost/GB) for user data.

    Would keeping user data on a conventional HDD affect these results? It would seem like it wouldn't, but I would like to see the evidence.

    I would really like to see more benchmarks for these drives that aren't synthetic. Have you tried things like Crysis or The Witcher load times? (Both seemed to me to have pretty slow loads for maps.) I don't know if these would be affected, but as real world applications, I think it makes sense to try them out.
    Reply
  • Anand Lal Shimpi - Monday, August 31, 2009 - link

    Personally I keep docs on my SSD but I keep pictures/music on a hard drive. Neither gets touched all that often in the grand scheme of things, but one is a lot smaller :)

    In The SSD Anthology I looked at Crysis load times. Performance didn't really improve when going to an SSD.

    Take care,
    Anand
    Reply
  • Eeqmcsq - Monday, August 31, 2009 - link

    I would have thought that the read speed of an SSD would have helped cut down some of the compile time. Is there any tool that lets you analyze disk usage vs cpu usage during the compile time, to see what percentage of the compile was spent reading/writing to disk vs CPU processing?

    Is there any way you can add a temperature test between an HDD and an SSD? I read a couple of Newegg reviews that say their SSDs got HOT after use, though I think that may have just been 1 particular brand that I don't remember. Also, there was at least one article online that tested an SSD vs an HDD and the SSD ran a little warmer than the HDD.

    Also, garbage collection does have one advantage: It's OS independent. I'm still using Ubuntu 8.04 at work, and I'm stuck on 8.04 because my development environment WORKS, and I won't risk upgrading and destabilizing it. A garbage collecting SSD would certainly be helpful for my system... though your compiling tests are now swaying me against an SSD upgrade. Doh!

    And just for fun, have you thought about running some of your benchmarks on a RAM drive? I'd like to see how far SSDs and SATA have to go before matching the speed of RAM.

    Finally, any word from JMicron and their supposed update to the much "loved" JMF602 controller? I'd like to see some non-stuttering cheapo SSDs enter the market and really bring the $$$/GB down, like the Kingston V-series. Also, I'd like to see a refresh in the PATA SSD market.

    "Am I relieved to be done with this article? You betcha." And I give you a great THANK YOU!!! for spending the time working on it. As usual, it was a great read.
    Reply
  • Per Hansson - Monday, August 31, 2009 - link

    Photofast have released Indilinx based PATA drives;
    http://www.photofastuk.com/engine/shop/category/G-...">http://www.photofastuk.com/engine/shop/category/G-...
    Reply
  • aggressor - Monday, August 31, 2009 - link

    What ever happened to the price drops that OCZ announced when the Intel G2 drives came out? I want 128GB for $280! Reply

Log in

Don't have an account? Sign up now