RAID Experiment: striping the outsides of platters = speedup?

Discussion in 'Mac Pro' started by JPamplin, Aug 7, 2009.

  1. JPamplin macrumors 6502

    JPamplin

    Joined:
    Mar 12, 2009
    Location:
    Nashville, TN
    #1
    Hi folks,

    I got bored with my setup (it's been the same for an entire WEEK, for Pete's sake), and decided to upgrade the 3 750s to 3 1TBs because the 1TBs were acting a little flaky on the nForce4 PC. We'll see how they do in the Pro.

    Anyway, I've decided to do a little test: I've partitioned the 1 TBs into a 150GB at the top and an 850GB at the bottom, then RAID0 the top 3 and the bottom 3 into 2 volumes - a 450GB for the OS/Apps, and a 2.55TB Media drive.

    Inspiration: http://www.tomshardware.com/reviews/short-stroking-hdd,2157.html

    I applied a 32k stripe size to the OS/Apps drive (perhaps I should have done 16K), and 64K to the Media drive. We'll see what the results are, throughput wise, but the random access numbers should go way up on the new OS/Apps raid because of the "outside o' the platter" partitions (TOP means OUTSIDE, right?)

    I'll post benchmarks when I get the OS installed and the Time Machine backup restored. Still running 10.5.8 - this is not an SL experiment.

    Can anyone comment with some degree of experience whether 16k or 32k is a better stripe size for an OS/Apps drive?

    JP
     
  2. nanofrog macrumors G4

    Joined:
    May 6, 2008
    #2
    Yes, Top = Outer. ;)

    Smaller stripe size will allow you to better utilize capacity (less "waste"), but it comes at a speed penalty, as there's more transactions to move a fixed amount of data. It will be mitigated a bit, due to the usage (OS), as they're small files and usage is random access.

    Give it a try, and see what you get. You can experiment with the stripe size, and dial it in that best suites OS X. ;)

    Might give you something to do this weekend, as you seem bored. :p
     
  3. gugucom macrumors 68020

    gugucom

    Joined:
    May 21, 2009
    Location:
    Munich, Germany
    #3
    I'll be back to this thread for results. How do you intend to do the short stroking? They used a special Hitatchi tool on all drives. Can you use that for other drives as well?
     
  4. seisend macrumors 6502a

    seisend

    Joined:
    Feb 20, 2009
    Location:
    Switzerland, ZG
    #4
    hah, I think this will take years for my setup to bore me :D

    I've read that the stripe size didn't made a lot of differences in some tests.
    But good luck, I would like to see your results. I selected the standart 32k.
     
  5. JPamplin thread starter macrumors 6502

    JPamplin

    Joined:
    Mar 12, 2009
    Location:
    Nashville, TN
    #5
    Holy Crap!

    OK, I'm certain this is not just the changeout in drives. Yes, I'm using Seagate 7200.12 1TBs instead of 7200.11 750s, but honestly this cannot be the entire reason.

    Observe:

    10.5.7 - RAID 1 (full partitions - 3 750s)

    Disk Test 103.12
    Sequential 150.78
    Uncached Write 211.75 130.01 MB/sec [4K blocks]
    Uncached Write 210.80 119.27 MB/sec [256K blocks]
    Uncached Read 75.37 22.06 MB/sec [4K blocks]
    Uncached Read 263.51 132.44 MB/sec [256K blocks]
    Random 78.35
    Uncached Write 33.68 3.57 MB/sec [4K blocks]
    Uncached Write 184.22 58.98 MB/sec [256K blocks]
    Uncached Read 112.74 0.80 MB/sec [4K blocks]
    Uncached Read 141.65 26.28 MB/sec [256K blocks]

    10.5.8 - RAID 2 (450GB outside partitions of 3 1TBs)


    Disk Test 199.76
    Sequential 252.34
    Uncached Write 439.41 269.79 MB/sec [4K blocks]
    Uncached Write 574.47 325.03 MB/sec [256K blocks]
    Uncached Read 96.73 28.31 MB/sec [4K blocks]
    Uncached Read 668.15 335.81 MB/sec [256K blocks]
    Random 165.31
    Uncached Write 60.01 6.35 MB/sec [4K blocks]
    Uncached Write 574.76 184.00 MB/sec [256K blocks]
    Uncached Read 350.17 2.48 MB/sec [4K blocks]
    Uncached Read 340.33 63.15 MB/sec [256K blocks]

    Look at these numbers! Wow!

    JP
     
  6. nanofrog macrumors G4

    Joined:
    May 6, 2008
    #6
    The biggest part is the 750's are the results of the entire drives, while the new configuration is only the fastest tracks. It makes a big difference on large drives. :) Not as much on smaller ones (i.e. 250GB models).

    If 2TB drives at 7200rpm show, those would knock the socks off most of us, in an experiment like this (just use the fastest tracks). ;)
     
  7. frimple macrumors 6502

    Joined:
    Nov 18, 2008
    Location:
    Denver, CO
    #7
    Well I would not have expected that... platter density playing a part is my guess.
     
  8. frimple macrumors 6502

    Joined:
    Nov 18, 2008
    Location:
    Denver, CO
    #8
    Huh, I always though the inside tracks were the fastest... I mean rotationally they're not sure, but I thought the density was the same across the platter making the outside slower because you had to seek across a greater distance.
     
  9. JPamplin thread starter macrumors 6502

    JPamplin

    Joined:
    Mar 12, 2009
    Location:
    Nashville, TN
    #9
    Think about it - the outer tracks contain more 0s and 1s per revolution than the inside tracks - for the same amount of time, the outer tracks transmit more data.

    I'm sold. Going with this setup from now on.

    JP
     
  10. gugucom macrumors 68020

    gugucom

    Joined:
    May 21, 2009
    Location:
    Munich, Germany
    #10
    You are going to use the inside partitions for mass data, are you?
     
  11. nanofrog macrumors G4

    Joined:
    May 6, 2008
    #11
    Yup. :)
     
  12. nanofrog macrumors G4

    Joined:
    May 6, 2008
    #12
    There's still about 1TB of capacity on the inner partitions though, before hitting the 50% mark, and getting bogged down by the inner tracks. ;)
     
  13. sammich macrumors 601

    sammich

    Joined:
    Sep 26, 2006
    Location:
    Sarcasmville.
    #13
    You almost had it. beaten to it, must be a slow morning for me

    Since you're rotating, the seek head covers more distance in one unit of time on the outside than on the inside.
     
  14. Shake 'n' Bake macrumors 68020

    Shake 'n' Bake

    Joined:
    Mar 2, 2009
    Location:
    Albany
    #14
    This is very interesting. How exactly do you accomplish this? I'm looking to get a RAID 1 or RAID 10 setup going in the family's XPS 410, and this would sweeten the deal.
     
  15. frimple macrumors 6502

    Joined:
    Nov 18, 2008
    Location:
    Denver, CO
    #15
    Yup yup! :) Makes perfect sense once you think about it... I gues I was envisioning radial spikes from the center containing 0's and 1's... :eek:
     
  16. JPamplin thread starter macrumors 6502

    JPamplin

    Joined:
    Mar 12, 2009
    Location:
    Nashville, TN
    #16
    @ frimple: yes, greater platter density is part of it, but not THAT much between a 750 and a 1TB. Using the outer tracks is most of the magic, I think.

    @gugucom: Yes, I just set up partitions for the remaining space of those 3 drives and RAIDed those into a biga$$ Media drive. Still use the space, just not for anything critical to performance.

    @shakenbake - match up your identical drives and wipe them for an OS install. Go into Disk Utility and create 2 partitions per disk, the top one should be small, and the remaining space is much larger - whatever you want, but the smaller you can make the top partitions without running out of room, the better.

    Then, go into the RAID section and create a striped array out of all 3 "1st" partitions. In my case, I created a 450GB boot volume out of 3 150GB 1st partitions of these 3 1TB drives, then the remaining space I RAIDed again into a Media volume. Install your OS on that boot volume, and enjoy, very simple.

    You guys going to try this? I would. This rocks.

    JP
     
  17. Shake 'n' Bake macrumors 68020

    Shake 'n' Bake

    Joined:
    Mar 2, 2009
    Location:
    Albany
    #17
    I'd do this in a heartbeat, but the only thing I've got running with multiple HDDs is my PowerMac server, and I don't think the G4 would be up to RAIDing. Plus the HDDs aren't the same capacity.
     
  18. gugucom macrumors 68020

    gugucom

    Joined:
    May 21, 2009
    Location:
    Munich, Germany
    #18
    Ooops got the G4 bit wrong.

    G5 would do it. The drives are SATA and would take 2x2TB. Time to upgrade?
     
  19. Shake 'n' Bake macrumors 68020

    Shake 'n' Bake

    Joined:
    Mar 2, 2009
    Location:
    Albany
    #19
    If I could get a G5, I'd love one, but they're still going for ~$1000 for a meaningful amount of power, although any G5 would be better than my 733 MHz G4.

    I am looking at an iMac to replace my Mac mini within the year, so If that does happen, the mini will be shifted to server duty and the G4 will be used mainly for HyperCard and light PS work.
     
  20. VirtualRain macrumors 603

    VirtualRain

    Joined:
    Aug 1, 2008
    Location:
    Vancouver, BC
    #20
    Correct me if I'm wrong but I think partitioning for short-stroking is largely a placebo effect. It looks good on benchmarks but real-world performance is the same whether you partition this way or not... Why I say this is because if you didn't bother and just striped all the drives as one large array/partition, then installed your OS then your Apps and then copied all your media over, you would effectively have the OS on the outer faster portion of the disk anyway and then the apps on the next fastest portion of the platters and so on.

    In fact, you could argue that partitioning it like you have for short-stroking is actually wasting some of the fast outer portion of the disk... you are forcing your media to a slower portion of the disk unnecessarily.

    As for stripe size, I've always thought the rule of thumb for OS arrays was 128K / # of disks in the array. Thus 64K for two disk arrays, 32K for four disk arrays, and so on. Thus you would be best off with 64K or 32K i think.
     
  21. gugucom macrumors 68020

    gugucom

    Joined:
    May 21, 2009
    Location:
    Munich, Germany
    #21
    VR, I'm nor so sure that your point would stand when the system gets older and apps get added with data. But I'm not at all a software expert and do not know how OS X does it's internal maintenance. Forcing the system to use certain fast path storage areas in an optimized way makes sense to me.

    If I had no SSDs I would probably go for three striped 1,5 TB drives and partition 5% from the top of them for boot/apps. That sounds like it could be mighty fast. One could still have a 2TB backup and a smaller Bootcamp drive that way.
     
  22. gugucom macrumors 68020

    gugucom

    Joined:
    May 21, 2009
    Location:
    Munich, Germany
    #22
    With meaningfull power you mean top end for G5 PMs? Where I live you can get a 2,3 GHz DP or DC for 450-700€. We tend to have used prices very simillar to $ prices in the US. So I would be surprised if you could not get a better G5 deal. Actually I guess we will see a lot of panic G5 sales with Snow Leopard and there may be some surprisingly good deals on a privat basis.
     
  23. JPamplin thread starter macrumors 6502

    JPamplin

    Joined:
    Mar 12, 2009
    Location:
    Nashville, TN
    #23
    Continuing with the experiment...

    Guys,

    I appreciate your comments on this, and here's what I'm going to do: I'm going to put the 750s back in a re-partition them as I have here with the 1TB, then run an XBench on that to do an apples-to-apples comparison.

    I've thought of one drawback to this setup, and a way to minimize it. If you really need to use the inner partition volume a lot, then the speed of your OS/Apps volume will decrease because the heads are jumping around. If that's the case, and you really need to use both volumes, then try SPANNING both volumes instead of STRIPING them - that way, the disks can act independently, so you'll get simultaneous throughput.

    The only problem with spanning is that you won't get the increased throughput from a RAID0 - you will, however, get the benefit of more throughput for the OS/Apps volume by forcing it to the outside.

    Will keep tinkering and let you know. Oh, and to answer the comment about speed being just a placebo - Leopard installed in just about half the time (like 30 minutes). I'll time it when I redo the 750s, but it was short.

    JP
     
  24. Tesselator macrumors 601

    Tesselator

    Joined:
    Jan 9, 2008
    Location:
    Japan
    #24
    Huh? No. That doesn't sound right. That's not how JBOD/Spanning/Concatenation works AFAIK.
     
  25. VirtualRain macrumors 603

    VirtualRain

    Joined:
    Aug 1, 2008
    Location:
    Vancouver, BC
    #25
    I believe spanning fills one drive and then moves onto the next. If so, it's actually the worst way to go for trying to keep stuff on the outer portion of the drive platters.

    As I said, creating a small outer platter partition for short-stroking only looks good on benchmarks. It offers no real-world benefits and as you've mentioned yourself, it can hurt seek times if you are accessing both partitions.

    The fastest simplest solution is to stripe your drives with a single partition, install your OS, update it, install your Apps, update them, then copy your media so that you have the most important software on the outer portions of your drives. Repeating this process from scratch once every 6 months to a year will ensure all your critical os and app software is on the "outer rim" :p

    Alternatively, having one drive for your OS/Apps and one for your media will ensure both are on the outer portion of their respective drives and you gain some parallelism.
     

Share This Page