Ideas for Real World HDD & SSD benchmarks ?

[wolf2009]

What REAL-WORLD benchmarks can one do, that can be easily be done by everybody else.

I was thinking copy some files from a drive to the same drive using this program
http://nodesoft.com/DiskBench/Default.aspx

Another benchmark could be muxing two 4 GB AVI files in Virtualdub.

What do you guys think ?

[mjz_5]

winrar.. copying 2 files at the same time, not just one.... bootup times.... program install times *use dameon tools)..

[extide]

Here is a pretty good real world example. Compiling Open Office. Usually takes 3+ hours. I did it in 35 mins with the SSD.

Installing Visual Studio 2008, usually takes ~25 mins. I did it in 3 with the SSD.

(I am comparing my 4x 160 GB Raid 0 Raptor stripe vs a single Falcon SSD)

[Megalomaniac]

you can follow what anandtech did, i think its the best example of how SSDs can be brought to their knees and gives real world examples

heres a page thats fun to read
http://www.anandtech.com/storage/showdoc.aspx?i=3531&p=13

they have a few more articles about these tests.Post your results after you test the drive :) then do the same on the HDD

[extide]

"used" performance isn't an issue for the Indilinx drives because of the 'wiper.exe' that is out there. It basically empties out all the "used but empty now" sectors on the drive so that they can be freshly written to later. I still get the same benchmarks on my drive as I did when i first got it about 6 weeks ago, and I have done a lot with the drive, including install several different OS's.

[AreEss]

The only effective sequential testing method is dd or dsmfmt with two clean drives. Anything else? Nope, it's not sequential. I've tested this far more than anyone else here, trust me on that. Source and destination must be matched and truly sequential operations. Otherwise, you skew results due to seek load on one side or the other.

Random is trickier, and honestly, the more common on desktop systems. Sequential benchmarks are completely meaningless for a gaming system, which is why I don't like things like HDTach and the like. They're all designed to make big numbers by doing sequential loading, with no regard for reality, because people want to see big numbers.

But, how do you do an actual random seek test? Short of bonnie++, I really have not found any effective method. You have no guarantee of fragmentation when loading games - especially Steam games, since it internally defrags - and no guarantee of the degree of fragmentation or randomization with anything else. In other words, you could do a test with a heavily fragmented large file, but you can't reproduce it. Which basically makes it useless, even same-system.
The "compile something" method for random? Completely useless; it's not truly random. It's CPU bound, not disk bound. The list goes on. There just is no reliable, reproducible method for random seek load testing.

[mikeblas]

It depends on what's real in your world, and what you're trying to measure.

Copying a file will largely show sequential burst I/O speed. This simply isn't important for my application; I'm more concerned with random I/O and latency times. To me, copying a couple files is pretty much idle. I run large database systems, and show my RAID array about 3000 I/O operations per second, with a mix of reads and writes, 24/7.

If you want a real world test, you should record what your real world access is like. There are various tools for this, depending on what your access application is and which OS you're using. Then, you can start replaying the workload against various devices, controllers, and configurations to see how you can influence performance with the different hardware and settings.

[extide]

You dont really need two drives to do a test with dd, you can use /dev/null as a destination (for read) and /dev/random as a source (for write)

There just is no reliable, reproducible method for random seek load testing.

Crystal disk mark's random 4k read/write benchmarks seem to be pretty reproducible, BUT I agree that they show what they show, not necessarily real-world performance.

(BUT if you have ever used a bunch of systems and ran that benchmark you WILL see a trend where general snappiness of the machine goes up with relation to those 4k random reads/writes.)