HachavBanav

Member
  • Content Count

    272
  • Joined

  • Last visited

Community Reputation

0 Neutral

About HachavBanav

  • Rank
    Member

Profile Information

  • Location
    Paris, France
  1. After some time spend on trying different solutions, I am sticking to those below: -(mainly) Boot on a 8GB DOM or the single local drive -Eventually use some RAM in a ramdrive (I use RamDisk+ that allow using hidden memory used by IO address) for temporary files -Run on a DAS : no raid (small pc) or raid 10 (reliable fast station) or raid 6 (file servers) -Backup everything online at BackBlaze (+Unlimited storage +NSA don't care +encryption +for $47.5/year) -Have some free cloud space for hot stuff (Currently I "DropBox") that require to synchronize files between nodes
  2. I would say your "2" choice is mine...but I would take care of using an "optimized" stripe size. I think most io size from VMs are 64Kb but if you have a database system or any app issuing updates on small io, please choose another storage target like raid 10
  3. To my understanding, all hw raid card reads a minimum of ONE stripe on ONE drive, even if it sends back only 4KB to the OS. Optimizing the stripe size is a VERY important point because it must be balanced between your needs of sequential io (which need the largest stripe size) and random io (which need a stripe size equals to the random io size). Very different applications requires TWO arrays : databases arrays may use small stripe (8-16-32KB) where backup arrays requires the largest one (1MB +)
  4. HachavBanav

    Adaptec 5805 - Raid-5 setup suggestions

    RAID controllers does not wait infinitely for a drive to answer. Usually they wait for up toi 7 seconds and consider the drive is dropped after this limit. Most "Desktop" drives has specific recovery error process to read a sector allowing for an almost-dead sector to be read but this recovery process may take up to 30 seconds. CCTL / TLER are the same feature that limits all drive's answers to stay below the 7 seconds timeout. Of course, all RAID edition drive has that feature. The point is that SOME tools may allow SOME drives to enable this feature. Without this feature, you shouldn't use a drive in a RAID array. WD is supposed to have a WDTLER.EXE tool to switch on/off this TLER feature...but I did not use it. Although many "desktop" drives has a very bad Unrecoverable Bit Error (UBE) ratio...down to "1 unreadable sector per 10^14bits read" which means you have about 8.8% of a sector error per 1TB read. Many many of those "RAID down" error you saw on the web are coming from this combination of a poor UBE and the lack of a timeout feature.
  5. HachavBanav

    Raid 5 HDtune performance

    Samsung HD204UI EcoGreen F4 : 2TB / 5400 rpm / Sata 300 / 4KB sectors / UBE at 1 per 10^15 Samsung HD203WI EcoGreen F3 : 2TB / 5400 rpm / Sata 300 / 512B sectors / UBE at 1 per 10^15 WD Caviar Green WD20EARS : 2TB / 5400 rpm / Sata 300 / 4KB sectors / UBE at 1 per 10^14 Can I suggest you to rebuild your array using 4 drives RAID 10 (+ 1 hotspare) instead of a RAID 5 without a large write cache backed by a battery ? Although, when partitionning your array volume, take care of aligning at a 64KB multiple (Vista, Win7 Win2008 aligns at 1MB where XP-Win2003 alignes at 31.5KB which would just KILL your performance)
  6. HachavBanav

    LVM plaid or MetaLUN?

    My point is that you have to take care of the io size pattern of all applications issuing IO on your array to optimize the stripe size. Maximum concurrency is reached (=IOPS is maxed) when 1 single IO requires 1 and only 1 stripe read/write. That way, one can expect to get up to 400 IOPS per SAS 15k drive using an io queue length of 8 (may be 16 or 32 will offer more IOPS but will increase rarely the maximum latency). So, doubling the stripe size (side effect of using LVM in your case) may not allow to deliver the expected IOPS
  7. HachavBanav

    LVM plaid or MetaLUN?

    Did you notice that replacing 6+6 15k SAS drives by TWO SSD in RAID 1 would have although answer your IOPS problem ? Appart from that, your maximum IOPS depends on the capability of setting your stripe size so that "1 query = 1 drive". ==> Depending on your current optimization, you have 2 situations: -current stripe is below your avg io size : doubling the stripe size will optimize IOPS for sure -current stripe equals (or more) your avg io size : doubling the stripe size will offer a negligeable raise in IOPS I agree that the choice between metalun and lvm does not really matters regarding performance but if the hosting guys feel better with one, I do agree with them as they will be faster to resolve downtimes and won't take bad decision under stress.
  8. Your raid card should have a scrubbing process that scans all blocks and check their readability and parity. I have no idea if 3ware have that and how it works for 3ware
  9. HachavBanav

    Recommended SSDs for RAID 5?

    0. Alternatives For "real" scratch drives (temp/swap), I would go RAID 0 or use some of your 24GB memory in a ramdrive For boot drive, I would use a small 100MB partition on your HDD For OS/Apps drive that the boot links to, I will create a SSD partition on a RAID 5 array taking care of having the RAID 5 stripe size to equals the OS "page" size (Windows cluster size can be up to 64KB) For scratch Data drives, I will use the OS/Apps partition above 1. No known problem if you know what aligning a partition is and if you have no useless background process (defrag, index, etc) 2. The TRIM function can not happen in RAID...and, in most case, you don't care at all because it only helps the garbage collector (GC) to know at deletion time whose sectors are to be resetted later on. In many situation, you only update sectors which means the SSD write on new sectors and marked to-be-resetted old ones. Obviously, if you leave enough free space (you already have 10% invisible "spare area" in all SSD) which is your case 3. SandForce 2000 series based SSD are the best as of today, Intel's are almost equivalent but don't worth the price gap to my opinion. Most SSD manufacturers provides TRIM alternatives (Intel's Toolbox for example) that "refresh" their SSD.
  10. HachavBanav

    6TB RAID 0 (dual 3TB drives)

    Striping drives should just MULTIPLY the sequential throughput (between 120 and 180MB/s for an hitachi 7K2000). I would say Windows striping has some overhead that cut about 10% of this throughput...
  11. Can you search for the current firmware as there is a BUG on HD204UI ? 40 drives with each 8.8% of an error per 1TB read means you HAVE errors...Starting the scrubber would be a very very good idea BEFORE a raid drive fail.
  12. Looks like Samsung retired from the HDD market... Anyway, the HD204UI Spinpoint F4 : -required an updated firmware -had a poor UBE of 1 unrecoverable sector per 10^14 bits read (8.8% of a read sector error per 1TB read) -had 4KB sectors (so partition alignment to 4KB was a mandatory challenge) Did you start the scrubber in your raid HBA ? This may fix the problem for you.
  13. HachavBanav

    RAID Card or NAS Box?

    Great news ! Just my 2 cents regarding 3TB drives: ">2TB" drives emulate 512-byte sectors and require a 4KB io size minimum + 4KB partition alignment (XP/Win2003 default partition reserved 31.5KB which does NOT align with 4KB; ext2 ext3 ext4 linux type fs although use the old cylinder-head-sector logic which usually leads to misaligned partition) Regarding Qnap vs Adaptec(in a black box) : Qnap will be easier to install/manage and I'll choose it if the price difference is not a real problem !
  14. No "System event" explains there are some IO errors ? What drives'model are you using on your raid 6 array ?
  15. HachavBanav

    RAID Card or NAS Box?

    RAID arrays are NOT backup solutions ! You may give a try to an online backup solution (like www.backblaze.com )