URE values - real or not?

When reading technical specification of disks given by vendors it can be noted that often makers provide not real Unrecoverable Error Rate values. This URE probability value is widely utilized to substantiate naive statements similar to "RAID 5 is dead by 2009" and to guess chances of double failure in RAID5. These calculations get people building their own RAIDs concerned.

In fact, the vendor URE data seems to be very far off the mark. Read technical documentation on Hitachi official website, they have kind of interesting URE values for 3 TB hard disk - 10^-14 errors per read bit. Such a value can be converted to the probability to read the drive from the start to the end and not encounter an URE is:

(1- 10^-14)^(8*3*1012)~0,79

therefore, the probability of the disk failing to read one sector is about 20%.

In other words when you have a disk filled at capacity there is a non-negligible chance (about 20%) that you are not be able to get data back off it. This is easily proven wrong by simplistic testing.

Random access time in RAIDs

The basic characteristics of a data storage speed are:

• access time which is defined as time delay between when a request is addressed to a storage device and the moment the requested data begins to come in. 
• throughput is sustained average transfer rate.

It is well-known that RAID level 0 increases throughput (N times increase ideally). When one plans to build a RAID array, only the throughput numbers are considered, not thinking about access time.

Access time in a regular disk includes the time to position a read head above the track (so called seek time) and the time which is needed a drive to bring a sector under a read head (so called rotational latency). No matter how many member disks are in RAID 0, there always exists such a sector which is simultaneously the furthest away from the head and this sector is not contained in the cache. For this sector the access time will be the same as in case of a single drive. The only way to decrease access time is to stick to an SSD.

P.S. One can easily estimate access time (and other performance characteristics) using free benchmark software BenchMe.

New free benchmark tool

I wanted to check my RAID from the performance point of view and so downloaded three benchmark utilities:

• HD Tune Pro
• Crystal Disk Mark
• BenchMe

 The price of HD Tune Pro is $34.95 and it provides linear read speed, access time, and quite a few other metrics most of which are not significant. I say these things because I just don't know where I could use this data. One more disadvantage is that it was difficult to find such a main performance characteristic as IOPS.

 CrystalDiskMark doesn't have an option to benchmark a physical data storage device, but is limited to partitions only. From my point of view this might be a drawback. The tool gives read/write speed for various number and size of requests. You can also get benchmark either for linear or random read/write speed.
All tests are designed in such a way that one is able to vary amount of data to be read.

And finally meet one more benchmark tool, BenchMe, that is a really easy-to use benchmark software. Unlike CrystalDiskMark, it benchmarks physical devices, but not logical volumes. Using the software one can get the various performance parameters such as linear read speed, distribution of access time, and a list of features. IOPS (I/O operations per second) values are measured for queue depth 1 and 32.

 Thus, the conclusions of benchmark software review are:

• HD Tune Pro is a paid software with a lot of useless benchmark parameters.
• CrystalDiskMark is a free tool which is characterized by inconvenient interface and with no ability to scan physical data storage devices.
• BenchMe is a free program that gives only really needed benchmark parameters in the most convenient representation.