One thing we have been working on for quite some time is getting a working model for doing simple MTTDL calculations on the web. After releasing reliability pieces over a year ago, several industry experts were quick to mention that they were not the best representations out there. There are more elegant models that one can use versus a simple Poisson distribution for sure, and it does not take into account other parts failing such as disk controllers, motherboards, power supplies and etc. It took a lot of back and forth but the basic idea is this, the calculator is “directionally” correct but is not the most accurate way to model all of that stuff. We did evaluate a much faster model but on an AWS m1.small instance it was taking over 15 minutes to complete with only one user. Simply put, this RAID reliability calculator will give you a fairly good idea regarding which RAID level is the most reliable given a number of drives. Also, if you do not believe manufacturer numbers, you can add things such as your own mean time between failure (MTBF) figures in this RAID reliability calculator and get meaningful output. The MTTDL model works fairly well for estimating this.
Some thoughts on setting values in the RAID reliability calculator:
1. Each year has approximately 8760 hours in it. You may want to adjust downward from manufacturer specs for real world use cases.
2. For real world usage, adjusting down a bit on the bit error rates works best.
3. For most newer disks 4KB sectors are the new standard over 512B sectors.
4. Volume rebuild speed is not necessarily a disks raw throughput (which, say on a modern 3.5″ SATA disk is 65-140MB/s.) Oftentimes this is negatively impacted by controller speeds, activity on the rest of the array and the access patterns during rebuilds.
If you find this useful, share the calculator! Tweet
If you see any issues or have suggestions, please post them on the official forum thread so we can fix them as soon as possible. Consider this a beta for now.