Wednesday, March 18, 2009

SSD followup

I wrote a couple of months ago how the Intel SSD's were the only ones on the market that seemed to be worth buying - all the cheaper ones were unusable due to having horrible random write performance, which is something you notice really quickly in real life as nasty pauses.

Sadly, almost none of the reviews seemed to ever catch on to that, as they were all looking at the (totally irrelevant) throughput numbers that basically don't matter in any real-life situation. Everybody just quoted the nice big marketing numbers, because finding the numbers that matter more to actual human perception (notably: average and maximum latency) was so much harder, and most disk benchmarks are crap and don't even give those numbers.

Which is why I was so happy to see this review at AnandTech. Half the numbers quoted are still the worthless ones (I guess you can't avoid quoting the industry standard benchmarks, even when they are horribly bad), but much of the actual discussion is about how unusable a drive is when it has maximum latencies in the hundreds of milliseconds.

Looks like people are catching on. And as the reviewers are catching on, so are the manufacturers. I still see too many reviews that gush over throughput numbers, but here's one that got it right, and apparently got a manufacturer to actually understand. Good job.

(And I still love my Intel SSD's , but Anand is certainly correct in pointing out that they aren't cheap. And it looks like "cheap" will no longer necessarily mean "sucks so bad that they are unusable" in the upcoming drives. Hallelujah)

26 comments:

  1. I'm pretty sure that what you refer to as "nasty pauses" are simple a feature that allows the user to ponder the mysteries of the universe (involuntary meditation). Really, it's not a bug, it's a feature.

    ReplyDelete
  2. So, have you noticed any of the now-famous slowdown/fragmentation effect on the X25-M?

    ReplyDelete
  3. The Register seems to be keen on the new Samsung SSD and reckons they'ved solved the stuttering problem whilst also being faster than the X25-M

    ReplyDelete
  4. I hope is that soon fast "small" (32GB) SLC becomes 'afordable' as they have a 'decent' number of delete cycle, combined with a huge HDD you have the best of both world:
    - put the OS, applications and configuration files and a dynamic cache in the SSD
    - put the multimedia (ie big) files on the HDD

    So I don't understand why SSD makers are so keen on improving their size for me a speedy SSD + a big HDD is a much better combination than a big SSD alone (well except for laptops).

    ReplyDelete
  5. Thomas B: I really haven't seen any major slowdowns. I'm sure they happen - GC gets more expensive as the drive is fragmented and fills up, but the original drive I got from Intel (and the one I did most of my benchmarking on) wasn't 'virgin' to begin with, so..

    Tom Fogg: I like the register, but their numbers there suck. It's all about those throughput numbers, they don't show any random-small-write performance numbers at all. Maybe Samsung got it right. No way to tell.

    ReplyDelete
  6. Lets all hope now that something like that happens to LCD market. Someone high profile like Linus needs to publicly scream at proliferation of crappy, glossy 6-bit displays with terrible viewing angles and pathetic color reproduction: only 262K colors.

    The sad part is: if you're shopping for a laptop, that 6-bit TN-film junk is all you can get. As far as I know not a single laptop manufacturer, not even Apple, offers 8-bit true color LCD panels on their laptops.

    ReplyDelete
  7. that's exactly what i'm always thinking of when i see those "reviews". it should be obligatory to publish the 4 values: read/write transfer, read/write access time.

    ReplyDelete
  8. what i would love to see is the SSD turning into something that can be used as a external archive media.

    ReplyDelete
  9. Anand is amazing. He is doing the million dollar work for SSD industry. I am happy that OCZ respected him and listened to him, now they have Vertex a potential winner.

    ReplyDelete
  10. I'm using the OCZ Vertex (30GB) as my root partition here (ext4 running gentoo).
    The drive is amazingly fast and stable so far. I did no tweaking except setting stripe-width=128 (512kb erase block size).
    No hangs, no stutter, cold$=hot$...

    Here are some iozone numbers:
    4kB random read and write ( File size set to 262144 KB):
    Conventional hdd: read:534, write:1370
    Vertex SSD: read:25412 write:14454

    ReplyDelete
  11. one more review:

    http://www.chip.de/artikel/SSD-Test-Extrem-Tuning-fuer-Notebook-und-PC-3_35143580.html

    the columns in the table at the bottom are:
    price, capacity, read, write, read (access time), write (access time)

    ReplyDelete
  12. This comment has been removed by a blog administrator.

    ReplyDelete
  13. I know it's not a manufacturer, but at least a major european hosting provider has been offering Intel SSD on its high-end servers for 6 month now.

    I'm not sure what keeps manufacturers at bay. Perhaps at this stage, Intel's industrial capacity is not up to their expectations, yet. I've read somewhere 10000 pieces/month, but I'm not sure the number is current.

    Anyway, the competition has apparently still not catched up...

    ReplyDelete
  14. Random write performance _is_ important. I notice that lots of reviewers think their benchmarking program is testing random write performance when it isn't.

    The ATTO Disk Benchmark, for example. It reports writes of 4KB at a time, but these are sequential writes, not random. It's actually very hard to find an SSD review that doesn't make this mistake.

    I tend to search for reviews that contain the term 'CrystalDiskMark'. At least that does test random write performance, if only within its 100MB allocated block.

    The best source of SSD reviews I've found is at:
    http://forum.ssdworld.ch/viewtopic.php?f=4&t=47
    The downside is that it isn't updated often. The h2benchw and CrystalDiskMark reviews that they do measure aspects of random write performance.

    ReplyDelete
  15. Thanks for the review nice observation.
    My Stupid Blog : earn money chao!

    ReplyDelete
  16. Allyn Malventano at http://www.pcper.com/article.php?aid=766&type=expert seems to know his stuff

    ReplyDelete
  17. mobile phones price and performance are a pair of contradiction. The word "shanzhai" is very popular in China. It is industry phenomenon of imitation, high quality and low price. Moreover, cell phones are the representative of them.The global financial crisis led to a significant shrinkage of assets for people. "china wholesale" demonstrate the might. most of them would like buy the cheap cell phones now .

    ReplyDelete
  18. Through google search engine, when you put in "discount cell phones", you will find more than 15,000 pieces of related information. Cell phone china has been concerned in the world. Many international sellers can get a considerable profit by ways of wholesale cell phones from China and switch selling. So it is a very good job on cell phones wholesale in the current financial crisis.Cellphone , it's not just talk tools.it make life easier.

    ReplyDelete
  19. I like this information about SSD followup, thanks for sharing!

    ReplyDelete