Technical details on the model » History » Version 1
Jessica Mack, 06/01/2015 08:20 PM
| 1 | 1 | Jessica Mack | h1. Technical details on the model |
|---|---|---|---|
| 2 | |||
| 3 | h4. Some background about replication and erasure coding |
||
| 4 | |||
| 5 | _"Erasure codes are a superset of replicated and RAID systems."_ |
||
| 6 | The main advantage of erasure codes is that they leverage the statistical stability of large number of components. |
||
| 7 | Assuming we have 1 million machines, and 10% are down we can calculate the probability of availability of a block that has 2 replicas using the formula below and we will get two-nines availability: |
||
| 8 | |||
| 9 | !Screen_Shot_2014-06-27_at_00.15.48.png! |
||
| 10 | |||
| 11 | Instead, we can use erasure coding with the same storage overhead ratio. For instance, we can use k=32, m=32 (in total we will have 64 blocks). The new formula is: |
||
| 12 | |||
| 13 | !Screen_Shot_2014-06-27_at_00.54.29.png! |
||
| 14 | |||
| 15 | This time with obtain over 8-nines availability. Nice, isn't it? |
||
| 16 | Source: Weatherspoon, H., Kubiatowicz, J., "Erasure Coding vs. Replication: A quantitative comparison" |
||
| 17 | |||
| 18 | h4. Rados |
||
| 19 | |||
| 20 | Both replications methods share assumptions: |
||
| 21 | * Parallel I/O recovery operations. |
||
| 22 | * Objects are written in the primary OSD of the PG identified by the CRUSH map. The primary daemon contacts other OSDs for replication and recovery purposes. |
||
| 23 | * Failure rate is constant and follows a Poisson distribution. |
||
| 24 | |||
| 25 | |||
| 26 | h4. State model (v 0.2) |
||
| 27 | |||
| 28 | !simple_state_model_v0.2.jpg! |