SSD Raid 0 vs separate drives

Discussion in 'SAMPLE Talk' started by Kevin Smithers, Jan 4, 2015.

  1. Kevin Smithers

    Kevin Smithers Senior Member

    46
    0
    Jul 10, 2013
    Hey guys,

    What do you think about this?
    I was wondering on which scenario I would get the best performance out of my system for sample streaming: having 2 1TB SSDs on RAID 0 or just splitting my libraries between the 2 drives with no RAID.

    I've read that RAID 0 would only benefit if I was reading big files rather than a lot of small files, like samples. Is this true? If not, what block size do you recommend?
    I would have a backup for all libraries so danger of loosing info on the disks is no issue, just looking for performance boost.

    I would be doing this with a Blackmagic Multidock (thunderbolt 2).

    Thanks!


    2014 Mac Pro
    OSX 10.9.5
     
  2. Siggi Mueller

    Siggi Mueller Senior Member

    228
    5
    Sep 27, 2009
    if you have a backup for all your libs, go for raid 0. in my studio with 2014 mac pro and maverick, i'm using 6 x 1tb ssd (samsung evo 840) for kontakt libs as raid 0 and 2 x 1tb ssd (samsung 850 pro) for Logic samples also as raid 0. i did benchmarking the two raid's and get a result of 800MB/s in reading. i choosed the highest block size 256kb. with one single 1tb ssd (samsung evo 840) i got about 500MB/s. Only the internal 1tb apple ssd in my mac is a bit faster (850MB/s) i think because it's hardwired as pci.

    make sure your ssd's are trim enabled.

    best, Siggi
     
  3. Wooloomooloo

    Wooloomooloo Senior Member

    59
    6
    Dec 31, 2014
    Brooklyn, NY
    In theory, if you're loading one thing at a time, then large files benefit from RAID 0 more often than smaller files, because they're more likely to be spread across the two drives. On the other hand, the large file will only load as quickly as the slowest drive can manage, assuming a 50/50 split of data (so RAID 0 will effectively make loading your files approximately twice as fast as your slowest drive alone could manage. If one drive is more than twice the speed of the other, which is unlikely, you won't get a benefit).

    In practice, when you're loading a large project and many samples at the same time, it doesn't matter because both drives are likely working as fast as they can anyway.

    RAID 0 has another advantage, which is that your backup can be of a single volume so replication is likely faster. If you just have two separate volumes, you'll have to back each up and manage the storage across them. Stick to Raid.
     
  4. JohnG

    JohnG Senior Member

    7,228
    3,196
    Nov 13, 2007
    California
    Interesting.

    I read that RAID really doesn't help much for the kind of small files we use (if we are talking about samples). I thought the main benefit was for larger files, like video.
     
  5. jsp21

    jsp21 Lurker

    35
    5
    Jan 4, 2015
    Raid 0 works by splitting data across multiple hard disks ("striping"), with the goal of maximizing file transfer speed. However, it doesn't reduce seek time and will most likely increase seek time considerably owing to the small size and large count of sample files. What Wooloomooloo is referring to is Raid 1, not Raid 0. Raid 1 works by mirroring the data in one hard disk onto another hard disk, with the goal of maximizing data integrity. If one hard disk fails, it can restore the data from the image.

    TLDR; don't go for Raid 0 configuration. You can either manually split your files across hard disks to speed up load times, or get SSDs.
     
  6. Wooloomooloo

    Wooloomooloo Senior Member

    59
    6
    Dec 31, 2014
    Brooklyn, NY
    I don't agree with this.

    SSDs and HDDs behave differently, so I would say that's largely true of HDDs (because the combined latency of the random reads would be greater than the added speed benefit). SSDs actually benefit far more from RAID 0 (not RAID 1) than HDDs ever did for this reason.

    Example, if each drives has a 15msec latency, 100 small files across two drives with identical performance could add as much as 3 seconds of searching. If you're only loading in a few hundred MBs of data, that could end up being slower on two drives than one. This would have certainly have been true 10 years ago when samples were relatively small anyway.

    With SSD latency in the 0.0xx Msecs range, you're always going to benefit from two drives rather than one assuming there are lots of small files, rather than just one or two, where the user will barely notice it.

    Now, with patches often being several GBs in size, even if made up of hundreds of smaller files, an SSD RAID 0 configuration should be a lot quicker than a single drive.
     
  7. JohnG

    JohnG Senior Member

    7,228
    3,196
    Nov 13, 2007
    California
    Thanks.

    Quite an interesting idea, especially as I've got SSDs on the vast majority of my samples now. And I fully appreciate the point about easier backups from a single volume.

    Not to be a Doubting Thomas, but have you done any tests with smaller file sizes? In general, I find that SSDs have allowed me to reduce my buffer susbstantially, but I would love to make it even tighter.

    I just wonder if the improvement is more in the "measurable" range but fails the all-important noticeable-when-composing" category.
     
  8. Wooloomooloo

    Wooloomooloo Senior Member

    59
    6
    Dec 31, 2014
    Brooklyn, NY
    I have a benchmark, for some reason I can't get the link to work.

    Results show a single SSD Vs RAID SSDs for 10GBs of files (one a single 10GB file, the other hundreds of files combined to make 10GBs of data). This is with an Intel 730 Series SSD

    Single 10GB file (SATA 6Gbps)

    Single SSD 376.2 Mb/sec
    RAID 0 707 Mb/sec

    10GBs of files and folders (like a sample library) SATA 6Gbps

    Single SSD 351 Mb/sec
    RAID 0 582 Mb/sec

    PCIe

    Single SSD 556 Mb/sec
    RAID 0 854 Mb/sec

    10GBs of files and folders

    Single SSD 465 Mb/sec
    RAID 0 677 Mb/sec

    So the benefit is best for huge files, but still you're talking about a 70% performance increase in the worst case scenario. So if a project takes 15 seconds to load on one SSD, you might get it down to 8 or 9 seconds. If you're at 10msec latency, perhaps down to 7 if your CPU is up to it.

    http://images.techhive.com/images/article/2014/06/plextor_ssd_raid-100315229-orig.png


    **edit - as an aside to clarify an earlier comment. RAID 0 always writes to both disks equally and in alternative stripes... so for there to be zero benefit from an SSD RAID 0, your file size would have to be smaller than the stripe size, which these days never happens. So I stand by what I said, which is that with SSD RAID 0 there is always a benefit, no matter how many and how small the files are. :D
     
  9. JohnG

    JohnG Senior Member

    7,228
    3,196
    Nov 13, 2007
    California
    Thank you -- as you argue, a very noticeable difference. Maybe you should consider publishing an article in Sound on Sound?

    But I still have one more question Do you have any thoughts or measurements about playback (voice count, pre-load size, other performance when using the samples)?

    Low on my list of worries is loading banks of samples. Whether it's 30 seconds or 2 minutes is, on the whole, pretty immaterial at least for me. Performance and low latency are key.

    I realise that I am pressing you, but my motivation is pure; this will help others on the forum and me too.
     
  10. mohurwitzmusic

    mohurwitzmusic Senior Member

    155
    5
    Jul 5, 2014
    Cool! Thanks for sharing this.

    Your data shows a great improvement in sequential read speeds, but
    do you also have test data about 4k Random Read? This would tell us more info about voice count.

    I also elected to forgo RAID because I got the same kind of advice as JohnG.
     
  11. Wooloomooloo

    Wooloomooloo Senior Member

    59
    6
    Dec 31, 2014
    Brooklyn, NY
    Good questions, no problem.

    Voice count will scale with the raw performance of the drives in a fairly linear fashion. So if you get 400 voices now, and a 50% speed increase from the drives being configured in RAID, then you'll get 600 voices more or less. Obviously this entirely depends on whether or not they're sequential or at some point in your project, 400 new voices kick in on one note!

    However latency will not decrease, because combining the dives in RAID 0 will not reduce seek/access time, in fact it will very slightly increase it (by a few nano-seconds, so not enough to be noticed by you).
     
  12. Wooloomooloo

    Wooloomooloo Senior Member

    59
    6
    Dec 31, 2014
    Brooklyn, NY

    OK, the answer can be a bit technical, but here is something to remember. There really is no reason to configure your RAID 0 configuration down to 4k blocks. You're right that 4k random reads will be slower on RAID 0 than a single drive, but only servers and synthetic benchmarks should care about this.

    I'd configure for 32k or even 64k. Now the argument is, large blocks waste more space. But in reality, the maximum amount of 'wasted' space is the block size multiplied by the number of files, and in reality you get no where near that amount of redundancy especially now that the average file size in your sample library is likely several hundred Kbytes or even larger.

    So if you have 10,000 samples, and you configure 32k blocks - the maximum theoretical space you'd waste would be 300Mb (I'm using base 10), but more likely much less than half of that. To waste the maximum amount, every single file would have to be exactly 1 byte larger than an integer number divisible by the block size (don't worry if that makes no sense, it's just extraordinarily unlikely).

    Most likely, on a 500GB drive, you might end up losing 150Mb or so by choosing 32k over 4k, and if that's really important to you, you'd have to weigh up performance vs space. 4k random is the worst possible scenario for any storage device, but I highly doubt even the worst file system of any musician is having to deal with 4k file fragments.
     
  13. mohurwitzmusic

    mohurwitzmusic Senior Member

    155
    5
    Jul 5, 2014
    Interesting! You are making a lot of sense.

    FWIW, I greatly appreciate you being as technical as possible :) There is a lot of hearsay on internet forums, so explaining it in technical terms is the best defense against this.

    Looks like I need to seriously reconsider RAID. Everything runs fine on my machines except Dimension Strings which is a resource hog of the worst kind. I really wish VSL would additionally save the DS ensemble patches as single files because streaming 8 patches at once to create a single note is slightly ridiculous. The .DAT files that accompany DS are .5 mb each.

    Best
    MOH
     
  14. Wooloomooloo

    Wooloomooloo Senior Member

    59
    6
    Dec 31, 2014
    Brooklyn, NY
    Thank you - I realize being new to a forum also adds a whiff of 'why should I believe you' as well, but this is actually a subject I am pretty comfortable with... just don't ask me about whether EQ should be applied before or after compression... I got into trouble answering that one somewhere else :shock:

    A lot of the common (mis)understandings of RAID (especially RAID 0) come from the days of low capacity mechanical drives. So compounding latency, reducing capacity to the limit of the smallest drive, poor interfaces (IDE anyone?) and some frankly woeful Windows drivers back in the day, just made it not worth it, unless you were a video editor. Also, people did generally format down to 4k to save space, which made the issue worse. There is a lot of misinformation on this in forums, but actually a lot of good literature online otherwise.

    SSDs are getting bigger, cheaper and the latencies are down to numbers much much smaller than typical audio latencies on a top end console - 0.08 msecs is typical. Before now, getting more RAM was far cheaper than getting a 2nd SSD, but once you're up to 64GB or so of RAM, doubling that or getting a 2nd SSD now favors the drive.

    It's a fairly recent turning point.
     
  15. zolhof

    zolhof Senior Member

    314
    363
    Jun 27, 2014
    Silent Hill
    Wooloomooloo, thanks for sharing your thoughts.

    Based on your experiences, if you were to install a couple of new 1tb SSDs (840 evo), in RAID 0, what would be your choices performance-wise:

    - GPT or MBR
    - alocation unit size (4k to 64k)

    Kind regards,
     
  16. OP
    OP
    Kevin Smithers

    Kevin Smithers Senior Member

    46
    0
    Jul 10, 2013
    Awesome, I guess I'll give RAID 0 a go.

    What is it to have my SSDs trim enabled?
    What about block size, Siggi suggested 256kb. Does everyone agree with that, Wooloomooloo?

    Thanks!
     
  17. mk282

    mk282 Senior Member

    1,865
    3
    Oct 25, 2011
    Personally I don't think RAID is really necessary for streaming a bunch of samples. Individual SSDs are plenty fast for that alone. You'll run out of CPU juice before getting DFD dropouts in any case.
     
  18. JohnG

    JohnG Senior Member

    7,228
    3,196
    Nov 13, 2007
    California
    I don't understand quite everything you are writing, but I appreciate the drift.

    Thank you!
     
  19. Wooloomooloo

    Wooloomooloo Senior Member

    59
    6
    Dec 31, 2014
    Brooklyn, NY
    MBR is fine up to 2.2TB, honestly there is really no difference between the two below that partition size, other than MBR supporting booting for WinXP and previous. MBR also still has more options for recovery tools if that's a consideration. If you go beyond 2TBs for a single partition, you have to use GPT even if it's across 2 or more drives.

    As for allocation size, larger stripes tend to offer better performance (on large files). Go for 64k if the majority of your files are that size or bigger. 32k if you have thousands of tiny samples smaller than 64k.

    Trim - all modern Operating Systems support Trim (I think after OS X 10.6 and after Windows 7).

    Block Size - 256k should be fine for a drive used for samples only, unless you have literally thousands of really tiny samples.
     
  20. Jason_D

    Jason_D Senior Member

    149
    29
    Jun 16, 2014
    California
    About a year ago I tried RAID 0 with two drives, then with three. There wasn't a huge difference like I thought there should have been in Kontakt load times. I am speculating that Kontakt has some sort of performance cap on loading capabilities. There's also scaling limitations as you increase the drive count. I wouldn't RAID 0 more than three drives.

    In benchmarks the load times were amazingly fast, around 1.2 GB/s.
     

Share This Page