Optane Memory - Is it for you? <Opinion/Review>

Discussion in 'SSDs & Data Storage' started by Keljian, May 31, 2018.

  1. Keljian

    Keljian Limp Gawd

    Messages:
    277
    Joined:
    Nov 7, 2006
    Over the last month or so, as a result of a thread on the Optane 900p, I have been experimenting with Optane memory and Primocache.

    What was established in the thread was that Optane has very low latency and high performance for low queue depth loads- which generally means that for desktop database applications (including things like browsers and games which use sqlite) there might be a benefit in having some Optane cache.

    I have not been sponsored by anyone - all costs for these tests have come out of my own pockets

    The system I'm running is:
    • Ryzen 1700, with a mild overclock (3.7 all cores)
    • Asus Prime x370 pro
    • 16 gig of ddr 3000 running at 14-14-14-34
    • Zotac AMP 1080ti (for completeness - not that it matters)

    In terms of what disks I have to test with - I have the following:
    • 950 -- 512gb Samsung 950 pro with cheap and cheerful heatsink from ebay
    • 850 -- 512gb Samsung 850 evo
    • MomentusXT - 512gb Seagate momentus XT (the original - with SD23 Firmware) - note updating the firmware on this has been a royal pain in the behind which I have spent hours trying to do without success over multiple machines, but it has been reliable.
    • Tosh128 -- 128gb Toshiba which I pulled out of a Dell Inspiron 11 3000 (THNSNK128GCS8)
    Software wise I'm using the Samsung NVME driver for the 950.

    As for the optane, it's a 32 gig original Optane memory drive, residing on a riser in the second 16x pci-e slot.

    Regarding configuration of primocache:
    • Blocksize is set to 4k
    • Ramcache is set to 1GB -as this is the minimum someone would likely use in a general use case
    • For both L1 (Ram) and L2 (optane) cache the cache is set to shared - so both Read and Write for the entire cache space
    • Deferred Write is set to on (10sec)
    • Prefetch is set to on
    • L2 size is set to Maximum
    FAQ
    Why haven't I used ATTO/IOMeter/HDTach?

    Because they're not relevant for the majority of people, and the sheer volume of data they produce with just four devices is obscene - they won't add anything relevant to the results

    What am I looking for?

    Something thing will justify to Joe/Jane Average that Optane Memory is worth it for them, over or with an NVME drive.

    Why do I care?

    Because the combination 32gig of optane + Primocache is a relatively inexpensive investment, it's probably in the range of a 256 gig SSD in terms of price - it's accessible. It's new technology, and I like playing with new technology

    Why am I doing this?

    Because no reviewer to date has put together a test like this. They've tested SATA SSDs and SATA HDDs but they have not:
    Run low blocksize cache (at block level)
    Tested with a reasonably fast NVME drive being cached
    Tested on a Ryzen system

    ....and few have tested with commodity SATA SSDs

    Will you Bench X/Y/Z?

    In short, yep, I'll keep adding to this thread as I go - suggest a test and I'll give it a whirl.



    On to the testing...

    CrystalDiskMark

    For CrystalDiskMark - I wanted to set it up such that the size of the dataset was larger than the ram cache - so we're focusing specifically on the optane. Thus I set the dataset size to 2GB from the original 1GB setting.

    CrystalmarkBaseline.JPG

    upload_2018-5-31_15-44-31.png

    So a few things become immediately clear

    • The MomentusXT speeds increase dramatically (as expected), as do the other SATA drives including the Samsung 850 evo
    • The 950's read speeds don't change in any great way at the top end, but beyond that, in the 4k range they improve dramatically
    • The 950's write speeds decrease in a big way at the top end (roughly a third of the performance of without cache).
    AS SSD Bench

    For the AS SSD Benchmark I did a similar thing - Set the working set size to 3GB, as 2gb isn't an option, so that it would saturate the 1GB ram cache. I also rearranged the results for readability, and ran the tests a few times each to make sure the results were consistent
    AS SSD Baseline.JPG AS SSD Results.JPG

    AS SSD Access times.JPG


    The things that stand out here:
    • Threaded 4k access - essentially a huge jump in performance for anything that uses a database.
    • Access times - don't quite double, but things get slower with cache direct access. I assume this is due to the fact you're adding some smarts to the storage layer.

    So in Conclusion
    SATA
    If you have a SATA drive as you primary data drive, no matter what it is, the combination of Optane and Primocache is worthwhile. You'll notice it when it comes to loading times, on windows boot times and you'll see a jump in write speeds.

    NVME
    On the surface, things are a little less clear with NVME. In addition to the tests I've run here, I've run several different cache sizes and the results are essentially the same. Write speeds go down. On the other hand you see read speeds, especially for 4k threaded loads, rocket up. The answer to what is better here depends on your use patterns.

    If you play games, web browse, and generally are a "light" storage user, then Optane will give your system a kick in the pants. If you are doing more heavy user type work where lots of files continuously need to be copied/written and read then stick to not using cache. That said I can't see an instance where you'd have this circumstance in a home environment.
    If you are doing this, it makes sense to limit the cache to 2-3gb or so, so that it fills quickly and overruns the data directly to the drive while still getting the benefit when you're using other desktop software.

    I guess this all says that even if you have NVME drives, it is worthwhile having the Optane/Primocache combo.
     
    Last edited: May 31, 2018
    craigdt and drescherjm like this.
  2. Keljian

    Keljian Limp Gawd

    Messages:
    277
    Joined:
    Nov 7, 2006
    The software is reasonably easy to tweak and you can cache multiple drives with one Optane (I have).

    upload_2018-6-1_13-4-38.png upload_2018-6-1_12-56-18.png upload_2018-6-1_12-58-19.png

    The key is the 4kb block size which plays to the Optane's strengths.

    Regarding setup. You have to set the L2 storage space up (format as NTFS, then allocate) and reduce the preset L1 cache size. I like defer write and am prepared to shoulder the risk as the L1 cache size isn't too big. Prefetching last cache makes things a bit quicker for booting
     
    Last edited: May 31, 2018
    JargonGR and craigdt like this.
  3. Meeho

    Meeho 2[H]4U

    Messages:
    3,887
    Joined:
    Aug 16, 2010
    Appreciate the effort, but this needs real life tests.
     
  4. Keljian

    Keljian Limp Gawd

    Messages:
    277
    Joined:
    Nov 7, 2006
    Give me a test to run, and I'll run it.
     
  5. westrock2000

    westrock2000 [H]ardForum Junkie

    Messages:
    10,351
    Joined:
    Jun 3, 2005
    I have the Momentus XT 750 and it is a good drive. It's too bad they never expanded on this in any meaniful way. The newer 3.5" drives still have a small amount of flash (8GB).

    I used the Sandisk Ready Cache 32GB on my work computer while it still had a spinner and that was a pretty good implementation of cacheing. Especially where I was able to daily see the difference of my computer opening programs versus my co-workers and their hard drive thrashing away.

    I do like the idea of caching. It would be great if the OS was aware and could optimize like Apple's Fusion Drive does (which also has no limit on SSD size). I don't know if it's really applicable to power users like much of us that have all SSD storage, but for regular people that want an OS drive with 1 - 2 terabytes, this is a great solution.
     
  6. Meeho

    Meeho 2[H]4U

    Messages:
    3,887
    Joined:
    Aug 16, 2010
    Programs/games load times and some of those benchmarks that simulate media editing/creation (PC benchmark?).
     
  7. Keljian

    Keljian Limp Gawd

    Messages:
    277
    Joined:
    Nov 7, 2006
  8. Keljian

    Keljian Limp Gawd

    Messages:
    277
    Joined:
    Nov 7, 2006
    Ok with 950 pro - FINAL FANTASY XIV: Stormblood Benchmark

    cache on 12.283 sec
    cache off 15.287 sec

    This correlates reasonably well with the results that Tweaktown managed, essentially showing that there's a 20% gain in performance with the optane caching the 950.
     
    Last edited: Jun 27, 2018
  9. Meeho

    Meeho 2[H]4U

    Messages:
    3,887
    Joined:
    Aug 16, 2010
    After which run?
     
  10. Archaea

    Archaea [H]ardForum Junkie

    Messages:
    8,435
    Joined:
    Oct 19, 2004
    I thought Optane didn’t run with Ryzen? That it needed a 7th gen Intel? What voodoo is this?
     
  11. Keljian

    Keljian Limp Gawd

    Messages:
    277
    Joined:
    Nov 7, 2006
    Run 2 (is it surprising that cache only works on subsequent runs? isn't that the idea?) - single run speeds are identical to no cache.

    Optane works with Ryzen, just acts like a normal disk
     
    Last edited: Jun 27, 2018
    drescherjm likes this.
  12. Meeho

    Meeho 2[H]4U

    Messages:
    3,887
    Joined:
    Aug 16, 2010
    It's not, but it is hard to be sure of the benefits if the cached data gets swapped often. And then there is the factor of how much of the game's assets are you loading multiple times and similar.
     
  13. Keljian

    Keljian Limp Gawd

    Messages:
    277
    Joined:
    Nov 7, 2006
    Yeah, that's why a 16/32 gig cache makes sense, you'd think it would swap less frequently
     
  14. drescherjm

    drescherjm [H]ardForum Junkie

    Messages:
    13,920
    Joined:
    Nov 19, 2008
    Intel's cache software does not run in Ryzen or older Intel boards. However you can use other software in windows and certainly in linux that provides similar (even better) features to the cache software from Intel.