Having just rebuilt my production desktop one week ago, I’m still in the process of tweaking and tuning that system to bring it up to max performance. Over the weekend, I installed Samsung Magician 4.4, the latest version of the SSD utility for the 500GB EVO 840 drive installed as the boot/system drive on that machine. Then, I ran a pair of tests to see what impact this had on system performance. By at least one measure, the difference is astounding, as the following before and after screenshots will attest:
BEFORE: CrystalDiskMark shows that the mSATA port on the Gigabyte Z77X-UD3H is running only at 3 Gbps (half-speed, in other words).
AFTER: CrystalDiskMark shows an improvement of 1.5 orders of magnitude, WEI shows no change. What gives?
Turns out that the massive performance boosts on sequential read and write shown in the first two blocks of CrystalDiskMark measurements while dramatic, simply don’t reflect much file system activity in the real world (except perhaps when transferring files larger than the 1 MB default block size shown). The second two blocks (4K normal access, and 4 K with a queue depth of 32) are closer to real life, except that queue depths on most Windows desktops seldom exceed a range of 6-8, even under heavy read/write IO loads (see Samsung’s informative discussion of “Benchmarking Utilities: What You Should Know” for more good information on what’s going on here).
Thus, the results that stay more or less the same for Sergey Tkachenko’s implementation of the old Windows Experience Index (WEI) for Windows 8.* (and the Windows 10 Technical Preview) really reflect the overall impact of the drive optimization software on performance for Windows desktops. That’s not to say that these utilities are worthless, or that you shouldn’t use them: firmware updates, disk optimization, and over-provisioning capabilities can indeed improve performance and extend the usable life of such devices. I just don’t think anybody should expect them to offer major performance improvements simply by virtue of their ongoing presence in the runtime environment. At best, I believe that improvements they can offer are incremental (probably less than 10% on overall I/O) rather than dramatic (an order of magnitude or better, as the first two blocks of the CrystalDiskMark results might lead one to hope for).