SSD results ATTO 16 and 32 KB read with Queue Depth 1?

Discussion in 'Memory and Storage' started by Alaind, Oct 30, 2012.

  1. Alaind

    Alaind Junior Member

    Joined:
    Oct 30, 2012
    Messages:
    12
    Likes Received:
    0
    Hi I'm looking for performance results for the ATTO benchmark without a Queue depth, aka checking neither on the right of the box. An application is using a memory mapped file and it gives 32KB sequential read operations. I didn't find those results in reviews, they all use QD4. I did it for my crucial M4 128 GB with a total length of 32MB and got those results : 16KB : 177592 32KB : 264274 64KB : 366351 Especially results for the samung 830 128GB and 256GB would be very welcome.
     
  2. Alaind

    Alaind Junior Member

    Joined:
    Oct 30, 2012
    Messages:
    12
    Likes Received:
    0
  3. razel

    razel Golden Member

    Joined:
    May 14, 2002
    Messages:
    1,756
    Likes Received:
    0
    You want a response within 2 hours in a pre-dominantly U.S. site during a time when most of us are asleep or waking up... all... for... free?
     
  4. jwilliams4200

    jwilliams4200 Senior member

    Joined:
    Apr 10, 2009
    Messages:
    532
    Likes Received:
    0
    You should use IOMeter rather than ATTO for accurate results. ATTO's results have always been rather weird, in my experience.
     
    #4 jwilliams4200, Oct 31, 2012
    Last edited: Oct 31, 2012
  5. Alaind

    Alaind Junior Member

    Joined:
    Oct 30, 2012
    Messages:
    12
    Likes Received:
    0
    maybe look at the day to.
     
  6. Alaind

    Alaind Junior Member

    Joined:
    Oct 30, 2012
    Messages:
    12
    Likes Received:
    0
    I use ATTO because Anand uses it in his bench. I'm quite sure he knows it better than me.
     
  7. jwilliams4200

    jwilliams4200 Senior member

    Joined:
    Apr 10, 2009
    Messages:
    532
    Likes Received:
    0
    Anand does NOT use ATTO with non-default settings like you are asking for.

    Also, Anand's reviews include IOMeter data.

    Besides, I guess that Anand would tell you the same thing I did, that IOMeter is more accurate and trustworthy than ATTO.
     
    #7 jwilliams4200, Oct 31, 2012
    Last edited: Nov 1, 2012
  8. Alaind

    Alaind Junior Member

    Joined:
    Oct 30, 2012
    Messages:
    12
    Likes Received:
    0
    Hi I reran the tests with IOmeter on my crucial M4 128. 32KB, 100% read, 100% sequential, 1 worker : first "run" : 247 MB/s second "run" : 244 MB/s The same test with 16 KB reads : 171 MB/s
     
  9. sub.mesa

    sub.mesa Senior member

    Joined:
    Feb 16, 2010
    Messages:
    611
    Likes Received:
    0
    You cannot *really* test sequential I/O with a queue depth of 1 unless you can disable the read ahead device cache and any read buffers that exist in software path.

    In plain English: the harddrive will transform a single queue depth with a predictable contiguous pattern into a multiple queue depth high enough to saturate its mechanical component. Yes, that is English.

    One important change exists between ATTO and IOmeter. ATTO tests via the filesystem and follow regular I/O API, only escapes filesystem buffering. IOmeter on the other hand is a raw I/O benchmark that totally circumvents the filesystem. This is called raw I/O and it means the filesystem will not perform read-ahead buffering for example, which turns your application single queue depth reads into high queue depth reads to your harddrive.

    If you test ATTO on a filesystem that doesn't start at a low offset or is filled with data, you could be testing the slower portions of your harddrive, as well. IOmeter can be configured to test the full LBA range.
     
  10. jwilliams4200

    jwilliams4200 Senior member

    Joined:
    Apr 10, 2009
    Messages:
    532
    Likes Received:
    0
  11. Alaind

    Alaind Junior Member

    Joined:
    Oct 30, 2012
    Messages:
    12
    Likes Received:
    0
    Unfortunally 4KB reads are far different than 32KB reads.
     
  12. Alaind

    Alaind Junior Member

    Joined:
    Oct 30, 2012
    Messages:
    12
    Likes Received:
    0
    I'm asking a simple question (for someone with a samsung 830) about a SSD not about a harddrive.
     
  13. jwilliams4200

    jwilliams4200 Senior member

    Joined:
    Apr 10, 2009
    Messages:
    532
    Likes Received:
    0
    Actually, they are 4KiB and 32KiB. And to be precise, they are 8 times different, which I would not say qualifies for "far different", which would surely have to be at least 10 times different.

    Anyway, if you want to compare posted 32KiB read data for various SSDs, then your best bet is to look for Anvil Storage Utility (ASU) results from reviews and forum posts. Although that is random, not sequential.

    Anandtech and Tom's hardware measures 128KiB sequential read data with IOMeter. You could try interpolating between 4KiB sequential read and 128KiB sequential read speed.

    I'm curious why you specify sequential. Wouldn't memory-mapped IO often be random? Or a mix of sequential and random?
     
    #13 jwilliams4200, Nov 2, 2012
    Last edited: Nov 2, 2012
  14. sub.mesa

    sub.mesa Senior member

    Joined:
    Feb 16, 2010
    Messages:
    611
    Likes Received:
    0
    What I'm saying is that you probably cannot test REAL sequential reads on an SSD without read ahead buffering. Only harddrives which allow their read ahead to be turned off will be capable of this kind of very rare performance testing, and it will show sequential read as under 1MB/s if done correctly.

    For SSDs, you can use IOmeter and other utilities to generate a single queue depth load on the device, but the SSD itself will use this predictable pattern and use read-ahead internally; thus still working with a higher queue depth.

    Without knowing what you want to test and why, I can't comment on it further. I just thought you wanted to see how sequential I/O performed without readahead or high queue depth.
     
  15. Alaind

    Alaind Junior Member

    Joined:
    Oct 30, 2012
    Messages:
    12
    Likes Received:
    0
    Well my testing shows that the application is reading 32KB at a time (QD1) and completely sequential for 37MB (in my case). My readings (and IOMeter and ATTO) for a Crucial M4 suggest that the SSD is not really optimizing for 32KB sequential reads with QD1. Looking at the Anandtech charts for ATTO I do see quite big differences how the different SSD react depending on the read size, but this with QD4, a situation where the SSD has less optimizing to do, I expect (and see on my M4) worse speeds with QD1. I'm thinking of buying a samsung 830 if that gives a real improvement. Someone with a samsung 830 and ATTO (or IOMeter) can run this test in less than a few minutes.
     
  16. jwilliams4200

    jwilliams4200 Senior member

    Joined:
    Apr 10, 2009
    Messages:
    532
    Likes Received:
    0
    I ran some IOMeter tests on a 128GB Samsung 830 and 128GB Plextor M5P (IOMeter 1.1.0-RC1 windows x86-64, Intel 6Gbps SATA ports, AHCI). I used an NTFS 8GiB file (16777216 sectors), one worker, 1 outstanding I/O per target, and did 100% sequential 100% reads (4KiB alignment) for various block sizes. The results are in MB/s (not MiB/s).

    block (KiB): 4, 8, 16, 32, 64, 128, 256, 512, 1024, 2048

    830 (MB/s): 78, 125, 220, 327, 394, 459, 485, 503, 503, 509

    M5P (MB/s): 68, 122, 205, 301, 368, 407, 407, 406, 407, 407

    By the way, I'd guess the Samsung 840 Pro might beat the 830 results, but probably not by a lot. Still, if you really need the highest speeds, you might want to consider waiting for the 840 Pro which is due out any day now.
     
    #16 jwilliams4200, Nov 3, 2012
    Last edited: Nov 3, 2012
  17. Alaind

    Alaind Junior Member

    Joined:
    Oct 30, 2012
    Messages:
    12
    Likes Received:
    0
    Thanks for the very useful numbers.
    The 840 pro is tempting, but a brand new type doesn't seems to fit in my "book".