[irq]: techie interrupted10/07/2012
“ Rebuilds tend to be expensive in any system preserving reliability. A 5 disk, fully-used RAID-5 volume with a failure needs to read 4 disks completely and write one disk completely to return to full protection and performance. This can take quite a while, and is getting worse over time as disk capacities increase faster than disk bandwidth. In an R=3 cluster, a single disk contains chunks of data replicated to other disks elsewhere in the cluster. For rebuild purposes, it’s nice to have at least 10 chunks on a disk, and it’s also nice for the 2 replicas for these 10 chunks to be on as many servers as possible. A rebuild then involves copying 1/10 of a disk between 10 pairs of systems, and the entire rebuild can be done in 1/10 the time of reading an entire disk. The rebuild of all the data on a failed server with 10 disks would involve 100 pairs of servers. While this eats up network bandwidth in comparison to any rebuild that takes place within a single server, the performance impact can be spread out over an entire cluster, instead of creating a performance hot-spot. While we’ve chosen 10 chunks per disk at blekko, some other NoSQL databases have as many as a million chunks per disk. Having a small number of chunks per disk leads to slow rebuilds that have big performance impact to a small number of nodes, and a performance hot-spot. Failing to spread out chunks properly also increases the time to rebuild and leaves hot-spots. „
High Scalability - High Scalability - Data Replication in NoSQLÂ Databases
Quote posted at 20:02 Comments
blog comments powered by Disqus