reparations

Member
  • Content Count

    250
  • Joined

  • Last visited

Community Reputation

0 Neutral

About reparations

  • Rank
    Member

Profile Information

  • Location
    southeast
  • Interests
    PC's, audio, cars
  1. reparations

    The Death Of Raid

    Bah, no edit, wrong witch :/
  2. reparations

    The Death Of Raid

    I've reserved comment on these proceedings because I knew there was little point to maintaining the truth in the face of a mob mentaility. But I can hold my tongue no longer after reading the crass comments in this thread towards anyone who might arrive at a different conclusion than one's own. This is all crap. There are many problems here that invalidate the claims that RAID 0 is pointless. First off is the test rig. Guess what, the impact of disk subsystem differences is blunted when you are cpu-bound. A 2GHz P4 may be a lot more respresntative than a P3-700, but neither is totally revealing in a world of 3.5-4GHz P4s and 2.2-2.5GHz FX-53s. Secondly, the testing regiment is crap. IPEAK results may be interesting and important in an indirect way, but they are not application testing. They attempt to model it, but are not. They prove no more than sandra drystones when it comes to comparing the effectiveness of different cpu architectures. They are the hard drive equivalent of an engine dyno, and any racer worth his salt knows you don't race dynos. The point where you arbitrarily choose what work pattern will represent this use or that use is the problem. Whatever pattern you choose is not the reality of the situation, and you have no way of quantifying if it is off or by how much for a given application. Thirdly is the choice of controllers. Guess what, if you are going to confine yourself with a 32bit 33MHz pci controller, STR can't help much. Pretty obvious if you ask me. This would point to the advisability of using a faster bus for connection to the system. And the cost effective solution? How about ICH5/R? Hmm? Perhaps the choice of the best RAID controller for desktop use might help the desktop-oriented results... hmm. Certainly RAID 0 degrades seek performance slightly. And that is a very important factor. But in all but the very worst case scenario the STR increase at least makes up for the loss, and in some areas allows a goodly effective increase in application performance. There is certainly the possiblitly of creating a laundry list of (cpu-limited, synthetic) benchmarks that show little advantage for it, but it is even easier to come up with a list of actual applications that do benefit. But to do so one must buy two cutting edge drives, use the correct controller (ICH5R for desktop use), and house it in fast enough machine to allow it be of benefit. Don't judge RAID0 by its inability to transform your KT333/1GHz Tbird system, properly integrated it into a modern, optimized system design and you can indeed benefit. And the data loss problem? The naysayers would be just as quick to nay at someone who says they can't be bothered to have implemented a proper backup regime, but then of course ignore that truth when it adds strength to the mob. And finally is the actual experience. If you round up the latest and best components and time application load times on RAID0 vs a single identical drive, the RAID box wins. Not hugely, but it wins. Generally 5-10%. IPEAK be damned. But I must congratulate you. The dogged pursuit of whiches real and imagined has had its impact. A fellow on my home board (ocforums.com) measured this real world performance--application load times--in a real machine (read of modern porportions): Let's compare the results: UT2004 14.32-12.12 = 2.2s (2.2s/14.32) x 100 = 15.4% improvement GC2 12.37-11.81 = .56s => 4.5% improvement Lock On 28.16-25.96 = 3.2s => 11.4% improvement Far Cry 90.42-83.70 = 6.72s => 7.5% improvement Unreal 2 16.44-15.90 = .54s => 3.3% improvement And even though he measured an 8.4% average improvement, immediately started chanting the anti-RAID0 mantra. I'll take my 8.4% and be happy, thank you, as this is the area least conducive to RAID0 improvement. It gets nothing but better from here.
  3. reparations

    The Death Of Raid

    LOL; whatever you do don't bring logic to a which-hunt
  4. reparations

    Building New system RE: Command Queing

    The point was though that it is not accurate to lump the performance of TCQ and NCQ into a single entity. TCQ is a pure loser for desktop use while NCQ is at worst a partial loser.
  5. reparations

    Building New system RE: Command Queing

    My understanding is that the above is true only for TCQ, it negatively impacting performance in a single user scenario through the increased load on the cpu. My understanding is that NCQ can offer substantial benefits, even in a single user scenario, as it is accomplished without loading the cpu to any great extent. As far as I know the new Seagate 7200.8 and the new Maxtor 16MB cache drives are the only ones that implement NCQ at this juncture, and the only testing I have seen of this capability used a prototype Promix TX4200 controller to enable the NCQ functionality. Results have looked positive, but understand that it is hard to implement NCQ today due to the lack of controller availability. The integrated drive controller on the new intel chipsets (i915/925) has NCQ support, but is at best a giant step sideways from the existing i865/875 chipsets in all other respects. Most will not have NCQ as an option until the release of the next round of disk controllers.
  6. reparations

    Looking for suggestions for my IDE dilemma

    Hindsight being 20-20, it probably would have been the time to go SATA on the latest hard drive and add a $20 Silicon Image SATA pci controller card. If you can return the drive I would still recommed this route. as the latest SATA Seagate version and all other makers' current SATA models outperform the seagate you have appreciably in a boot-drive scenario. If you don't mind the comparitvely lazy performance, Hook the seagate up to your Promise. I have a Promise Ultra 100TX2, both it and the Ultra 100 support 48bit LBA (for drives larger than 137GB), but only if flashed with the latest bioses. Optical drives on a promise are always very sketchy. May work, may not. I'd run the optical off the mb and the hard drives off the Promise for sure. I'd put the 120WD on the other channel of the Promise.
  7. reparations

    Building New system RE: Command Queing

    My understanding is that the Raptor supports TCQ, not NCQ. TCQ has proven to slow single user desktop usage patterns, only helping under server usage. WD's intent is to use TCQ to allow the Raptor to more fully compete with SCSI drives for server use, but it is preferable for your application to leave ot disabled.
  8. reparations

    Raid 0

    Anytime you are building a pc primarily for video or sound editing and/or capture a strong disk subsystem is a good idea. In years past that mandated using a SCSI drive or an IDE RAID0 array. Single IDE drives weren't fast enough on the inner tracks to cut the mustard. Nowdays the STR of the better IDE (or SATA, same dif) drives exceeds 30MB/s. Single drives are successful in the majority of applications. But if you have a robust backup plan an RAID0 array is still good insurance against running out of STR (which can happen more readily than the raw numbers might suggest) without introducing significant cost. SCSI RAID, while certainly nice, is a bit of overkill, but a single fast SCSI drive might be just the ticket if you can deal with their inherntly higher noise floors as compared to even two modern IDE/SATA drives.
  9. reparations

    Photoshop Cs batching machine

    Which kind of points to the fact that the Opteron is more powerful, giving competitive performance with the handicap of the 32bit version of the application. At least I'm assuming the '64bit' in Apple land really is an advantage. In the end it seems as though macs are inherently less powerful in architecture (or just lack clock speed), especially at any sane price point.
  10. reparations

    Lousy RAID performance

    Don't endevor to estimate real world performance, measure it. A stopwatch is your friend. Measure things like time to load windows, note control panel loading (should be pretty damn fast), and install your slowest-loading application and time it. In addition you may wish to capture audio and/or video or rip/burn DVD's to insure your STR is adequate and perform a few simple file copy/delete tests with your own data. Think about the things that take time, and time them.
  11. reparations

    Lousy RAID performance

    I figured the HDTach anomoly resulted from the 64K stripe. As far as what stripe is best for you, a little stopwatch work on pertinant tasks is the best route. There is no guarantee that a 64k stripe is not better, but I certainly didn't notice any improvement on my 7k250 machine as compared to its performance when configured with the 16k stripe. Most recent testing I have seen has indicated 64k as being attractive, however. I certainly haven't ever encountered an audible differnce result from choice of stripe size. I had intended to obtain the WB99 disk inpection test and see if the STR anomoly is present in its result as well. HDTach is far from the final word on anything, but you did ask where the effect comes from
  12. reparations

    Lousy RAID performance

    I have historically used 16k stripes for my RAID O stuff. Smaller stripes favor random access, where large ones favor large record moves. Since STR is rarely a limitation with RAID 0 setups I always tried to minimize the random access penalties by using the 16k stripe. I saw some testing that claimed rather striking benefits for a 64k stripe, and that is why I tried it. If I still had the drives I would go back to 16k. I assume you did insure that file indexing is turned off. If you have a spare HD you can ghost your existing XP install off, re-configure the array, and spit it back on. Lots more fun than a re-install. Ghost 7.5 and up will work with your sata controller, or Acronis TrueImage.
  13. reparations

    Photoshop Cs batching machine

    You might also try overclocking the thing; my 2.8c runs 3.5GHz rather effortlessly. When you live and die by the clock cycle, oc'ing makes a fair degree of sense. Another 700MHz out of 2800 will have a basically porportional effect on your processing times.
  14. reparations

    Lousy RAID performance

    Ahh, that would make sense. But it is strange it only affects the arrays. The single drives in the system read normally...
  15. reparations

    Lousy RAID performance

    I noticed the same thing when I laid down a fresh intall of XP SP1 on a machine with 2x80GB Hitachi 7k250's on an ICH5R controller. The only things I did different on that install versus the previous one (where the STR graphs looked normal) was to try the 64KB strip size (previously had 16) and use the latest version of the Intel ICH5R drivers rather than the older ones I had been using. The drives have since been sold, and I did not have time to figure out which of the two at fault, but you might try a 16 or 32k stripe just to see.