• Register
  • Help
Results 1 to 8 of 8

Topic: [vcache] revealed

Share/Bookmark
  1. #1

    [vcache] revealed

    After much experimentation with a wonderful little program called DSKBENCH, which pinpoints hard drive performance as it relates to audio streaming, here are my optimum settings for excellent Giga performance:

    [vcache]
    ChunkSize=1024
    MinFileCache=16384
    MaxFileCache=16384

    Here\'s the disk performance numbers for the above settings on a Maxtor DiamondMaxPlus 40GB UDMA IDE HDD:

    DskBench 2.11
    (c) 1998, SESA, J.M.Catena (admin@sesa.es, www.sesa.es)
    Timer Check = 991 (should be near 1000)
    CPU Check = 50.09 % (should be near 50.00 %)
    CPU index (relative to Pro 200 MHz) = 3.573466
    Open = 1 ms
    Write = 12359 ms, 20.71 MB/s, CPU = 1.98 %
    Flush = 16 ms
    Rewin = 0 ms
    Read = 9000 ms, 28.44 MB/s, CPU = 2.49 %
    Close = 20 ms
    BlockSize = 131072, MB/s = 10.99, Tracks = 130.62, CPU = 2.42 %
    BlockSize = 65536, MB/s = 6.60, Tracks = 78.46, CPU = 1.21 %
    BlockSize = 32768, MB/s = 4.15, Tracks = 49.29, CPU = 1.22 %
    BlockSize = 16384, MB/s = 3.93, Tracks = 46.68, CPU = 1.77 %
    BlockSize = 8192, MB/s = 3.88, Tracks = 46.16, CPU = 3.01 %
    BlockSize = 4096, MB/s = 3.89, Tracks = 46.27, CPU = 5.58 %

    The key setting is not MaxFileCache or MinFileCache, but rather ChunkSize. Check out how the CPU usage jumps to 3 or 4 times higher when ChunkSize is disabled:

    DskBench 2.11
    (c) 1998, SESA, J.M.Catena (admin@sesa.es, www.sesa.es)
    Timer Check = 990 (should be near 1000)
    CPU Check = 52.46 % (should be near 50.00 %)
    CPU index (relative to Pro 200 MHz) = 3.400629
    Open = 1 ms
    Write = 12189 ms, 21.00 MB/s, CPU = 6.60 %
    Flush = 14 ms
    Rewin = 0 ms
    Read = 8738 ms, 29.30 MB/s, CPU = 7.53 %
    Close = 17 ms
    BlockSize = 131072, MB/s = 11.05, Tracks = 131.38, CPU = 8.23 %
    BlockSize = 65536, MB/s = 6.66, Tracks = 79.23, CPU = 9.10 %
    BlockSize = 32768, MB/s = 4.04, Tracks = 48.06, CPU = 9.34 %
    BlockSize = 16384, MB/s = 2.65, Tracks = 31.46, CPU = 9.81 %
    BlockSize = 8192, MB/s = 2.55, Tracks = 30.35, CPU = 11.90 %
    BlockSize = 4096, MB/s = 2.40, Tracks = 28.52, CPU = 13.70 %

    There go all your voices >

    Larger MaxFileCache sizes tend to not improve Sustained Throughput at all, they actually stay about the same as CPU usage on Read increases by about 50%.

    I\'ve also tested sustained throughput on my 75GB Seagate Cheetah and 3 other Maxtor IDE drives and the increase or decrease in performance is consistent.

    You can find DSKBENCH at ProRec.com

    Happy tweaking!

    Lennie Moore
    lamoore@loop.com

  2. #2

    Re: [vcache] revealed

    Lennie,

    Thanks for the tip. I tried it, but don\'t really know how to interprete the results. Also how did you manage to print out the result? For me, the DOS prompt window just disappears when the test is finished. Maybe because I open a DOS prompt from windows?

    I wrote down these results before it shut down:

    Write 19883ms 12.88MB/s cpu 1.87%
    Read 11180ms 22.90Mb/s cpu 2.45%

    I have Maxtor 7,200rpm drives on ATA33. Would my results indicate a bottleneck preventing full 160 voices polyphony?

    What do you think?

    Hans

    [This message has been edited by Hans Adamson (edited 03-21-2001).]

  3. #3

    Re: [vcache] revealed

    To write the results to a file, do the following:

    1. Copy the DSKBENCH.EXE file into C:WINDOWS
    2. Go to an MS-DOS prompt (Start>Programs>Accessories>MS-DOS Prompt in Win98) and type n:<Enter> where n=the drive letter you want to test.

    At the drive prompt type
    dskbench>n:results.txt<Enter>
    and the program will write the results to a txt file.

    Don\'t worry if it looks like nothing\'s going on, the program is working. This test take 5-10 minutes so grab a cup-o-joe, wait a few minutes and when you see the drive prompt again with a flashing cursor, the program has completed its process and your txt file is ready.

    As far as interpreting the results, we\'ll use yours as an example:

    <BLOCKQUOTE><font size=\"1\" face=\"Verdana, Arial\">quote:</font><HR>
    I wrote down these results before it shut down:
    Write 19883ms 12.88MB/s cpu 1.87%
    Read 11180ms 22.90Mb/s cpu 2.45%

    I have Maxtor 7,200rpm drives on ATA33.
    <HR></BLOCKQUOTE>

    11180ms is how long it took your hard drive to read an uncached 256MB file.

    22.90MB/s is your hard drives\' true sustained transfer rate. This is one of the important numbers as it tells you how many MBs per second your hard drive can read on a continuous basis.

    CPU 2.45% is how much of the processor\'s time needed to be spent on setting up the transfers. This is the other important number as it tells you how much you\'re taxing your processor when you\'re reading from your hard drive at the sustained transfer rate.

    Your numbers look great for an IDE drive. The Maxtors tend to have some of the best disk performance numbers.

    <BLOCKQUOTE><font size=\"1\" face=\"Verdana, Arial\">quote:</font><HR>
    Would my results indicate a bottleneck preventing full 160 voices polyphony?
    <HR></BLOCKQUOTE>

    No, they look excellent. Your hardware seems fine. My guess is there may be something hiding in your registry that causes Gigastudio to only perform with less than 160 voices. Either something weird happened with Windows, which wouldn\'t surprise me, or with Gigastudio\'s registry entries, which would be Windows fault, which wouldn\'t surprise me. Sorry I don\'t have a more complete solution for you. If you want assistance with Windows tweaks or registry surgery relating to gigastudio, e-mail me and I\'ll try to point you in a good direction.

    Best regards,


    ------------------
    Lennie Moore
    lamoore@loop.com
    http://members.loop.com/~lamoore

  4. #4

    Re: [vcache] revealed

    Lennie

    Nice post!!!

    How does that equate to your system in real terms??? For example, can acheive heigher polyphony, greater sequencer track counts...etc.

    TIA.

    Brian.

  5. #5

    Re: [vcache] revealed

    I imagine we\'d see a significant difference between MB chipsets. It\'s fairly apparent that some VIA chipsets don\'t perform memory functions as efficiently or as well as Intel chipsets, KT133 series being an exception. I\'d be curious to see the data comparing different hdw platforms.

    Steve Mitchell
    The Classical MIDI Resource
    The CMR Players

  6. #6

    Re: [vcache] revealed

    Track counts are listed in the above results under the various BlockSizes tested. They refer to how many mono 44.1kHz tracks can play at the same time with the respective filesizes.

    For example, the first BlockSize is 128K and the track count is 130-ish mono tracks at 44.1kHz. The lower Track counts for smaller BlockSizes make sense as your hard drives\' read/write heads have to start and stop and relocate and all this work with smaller BlockSizes takes more time than just reading from one large track like we have in Gigasampler files and hard disc recording.

    This all applies to audio track performance within programs such as Logic Audio Platinum or Cakewalk Pro Audio, although how well each program handles audio streaming efficiently is a matter of whether they did a good job in their respective programming and design.

    As far as a specific way to translate the results here, all I can say is I need more information from Nemesys. What are they basing their 160 voices on? 20MB/s sustained throughput divided by 160 equals 128KB/s for each voice to work. Does a stereo wave file count as two voices? I would think so. What about Layers? How do they affect voice count?

    I guess my point is that there are too many unanswered questions. I don\'t push my system with Performances that exceed 70%. I only load what I need and if I came up to a wall where I was maxing out my Gigastudio, I would build a second system and share the workload between the two. By itself it still kicks *** over a single maxed out Roland S-760 any day. I think in general people expect too much. My experience is that PCs still have a ways to go and even then we won\'t be happy \'cause we want them to read our minds, cover up for our own errors, and make us breakfast, too.

    Let\'s keep demanding excellence from the Nemesys folks as well as the hardware manufacturers and have a great time creating with all these cool new toys.

    Best regards,


    ------------------
    Lennie Moore
    lamoore@loop.com
    www.lenniemoore.com

  7. #7

    Re: [vcache] revealed

    <BLOCKQUOTE><font size=\"1\" face=\"Verdana, Arial\">quote:</font><HR>Originally posted by SteveMitchell:
    I imagine we\'d see a significant difference between MB chipsets. It\'s fairly apparent that some VIA chipsets don\'t perform memory functions as efficiently or as well as Intel chipsets, KT133 series being an exception. I\'d be curious to see the data comparing different hdw platforms.

    Steve Mitchell
    <HR></BLOCKQUOTE>

    Steve,

    I agree that there may be a little hit in performance between a motherboard with a 440BX chipset and a VIA chipset. I\'m not sure if it\'s that substantial.

    Check these test results out:

    40GB Maxtor DiamondMax Plus 7200 rpm UATA66 HDD on my Gigastudio computer with a Tyan Trinity 400 (S1854) motherboard with a VIA Apollo Pro chipset...

    DskBench 2.11
    (c) 1998, SESA, J.M.Catena (admin@sesa.es, www.sesa.es)
    Timer Check = 988 (should be near 1000)
    CPU Check = 50.10 % (should be near 50.00 %)
    CPU index (relative to Pro 200 MHz) = 3.212552
    Open = 14 ms
    Write = 9742 ms, 26.28 MB/s, CPU = 3.75 %
    Flush = 19 ms
    Rewin = 0 ms
    Read = 9266 ms, 27.63 MB/s, CPU = 3.23 %
    Close = 40 ms
    BlockSize = 131072, MB/s = 10.86, Tracks = 129.14, CPU = 2.25 %
    BlockSize = 65536, MB/s = 7.27, Tracks = 86.42, CPU = 1.29 %
    BlockSize = 32768, MB/s = 3.77, Tracks = 44.85, CPU = 1.10 %
    BlockSize = 16384, MB/s = 1.93, Tracks = 22.93, CPU = 1.07 %
    BlockSize = 8192, MB/s = 3.00, Tracks = 35.71, CPU = 2.50 %
    BlockSize = 4096, MB/s = 5.21, Tracks = 61.98, CPU = 7.66 %

    And the same 40GB Maxtor 7200rpm UATA66 HDD on my sequencing computer with an ABIT BX133-RAID motherboard and the Intel®440BX chipset (82443BX and 82371EB)

    DskBench 2.11
    (c) 1998, SESA, J.M.Catena (admin@sesa.es, www.sesa.es)
    Timer Check = 990 (should be near 1000)
    CPU Check = 52.04 % (should be near 50.00 %)
    CPU index (relative to Pro 200 MHz) = 3.565500
    Open = 1 ms
    Write = 12543 ms, 20.41 MB/s, CPU = 1.87 %
    Flush = 17 ms
    Rewin = 0 ms
    Read = 8816 ms, 29.04 MB/s, CPU = 3.23 %
    Close = 25 ms
    BlockSize = 131072, MB/s = 10.92, Tracks = 129.83, CPU = 3.13 %
    BlockSize = 65536, MB/s = 6.63, Tracks = 78.86, CPU = 1.25 %
    BlockSize = 32768, MB/s = 4.13, Tracks = 49.08, CPU = 1.24 %
    BlockSize = 16384, MB/s = 4.04, Tracks = 47.99, CPU = 1.89 %
    BlockSize = 8192, MB/s = 3.99, Tracks = 47.43, CPU = 3.31 %
    BlockSize = 4096, MB/s = 3.96, Tracks = 47.02, CPU = 5.98 %

    Pretty close, I\'d say. At least close enough on the Read results to keep the sustained throughput, CPU%, and the Track counts similar.

    It\'s an interesting question. I\'ll test it out more fully down the road. Today I need to get back to writing music

    Regards,



    ------------------
    Lennie Moore
    lamoore@loop.com
    www.lenniemoore.com

  8. #8

    Re: [vcache] revealed

    There are only two tiny problems with this:

    a) That program puts significant stress on a harddrive. The program comes with a warning - take it seriously, I have already lost one (Western Digital) harddrive to this.

    b) The sustained throughput rate has no importance whatsoever for Gigastudio performance, which is why you are battling to equate 20MB/s to 160 voices polyphony.

    What is important, is the random block read rate. For a typical harddrive with access time around 5ms (close to drive start), this can be measured with something like Sisoft sandra lite to be something around 4MB/s. (Which equates nicely to 160 voices).

    This is directly related to the access time of the drive. As I have said numerous times before, some plug-in cards add around 20ms latency, which makes the random block read rate go to like 1MB/s, with a corresponding drop in achievable polyphony. One can calculate this and measure it with Sisoft and Gst - real world performance seems to correspond to within 15% or so of what is expected.

    The reason that this is the important parameter, is simply that GSt reads a bit of one sample, then reads a bit of the next sample, etc. - so the heads are jumping around the whole time.

    Also, what will have a great effect on HD performance, is the old story of not having DMA enabled.

    As far as I can remember, the Connexant patent pre-caches samples by 64kB. So I would guess that if you could set your chunksize to 16k, 32k or 64k, it might be usefull.

    Yes, stereo samples do indeed use extra voices, as do layers within a sample - play a note and see how many voices Gst indicates that it is using.

    PS: The last time I checked, the memory indicator in Gst does not reflect memory allocated to vcache. So, if you make your vcache equal 25% of your memory, your PC will start behaving very strangely once you have loaded the Gst memory indicator past about 73% or so.


    [This message has been edited by cc (edited 03-23-2001).]

Go Back to forum

Tags for this Thread

Bookmarks

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •