|
Cache is another important pre-purchase consideration because it can have a major impact on I/O performance. Highly sequential workloads require lots of cache because that's where most of the data will be read. Cache read-ahead algorithms try to predict I/O requests by pre-staging data into cache. However, cache is a zero-sum game; when something goes in, something else gets bumped out. The more cache you have, the more pre-staged data is maintained and performance is improved by not having to fetch data from disk. But cache memory isn't cheap and should be increased only if the application requires a larger cache to run properly.
High write activity, for example, may require more cache because with some controllers write activity may be mirrored between clustered controllers and cached before being written back to disk. These options will be discussed more later. Before purchasing an array, you must determine if you'll use write-back caching and/or write mirroring, and then size your cache accordingly. In addition, on some high-end subsystems, asynchronous remote mirroring of I/O activity consumes additional cache. On midrange systems, remote replication takes less cache, as data and meta data are typically held on disk only while being replicated. Sometimes onboard replication of data (point-intime copies) uses cache to retain updates while the copy is active.
Configuring LUNs and RAID groups
Most I/O subsystems map LUNs to specific RAID groups, and the RAID level you use to support those LUNs affects performance. Also, in most midrange systems LUNs are typically assigned to a preferred controller in a cluster of controllers, although this can change during runtime. If there are LUNs reserved for specific controllers, dividing LUN I/O activity across multiple controllers can balance the controller workload.
Most subsystems support several RAID levels. RAID 5 and RAID 1 are typical, and RAID 5+0 (or 50) and 1+0 (10) may also be available. The choice of RAID level is usually based on the workload's read-to-write ratio and randomness. Although the write penalty for RAID 5 keeps getting smaller, it still exists. For most highly random, write-heavy workloads, it's probably wise to use RAID 1. For sequential workloads with low or high write activity, RAID 5 will suffice because sequential writes have less of a write penalty than random I/O. Throughput isn't a big factor when choosing a RAID level as long as the stripe can sustain the workload. However, cost is a factor to consider: RAID 1 consumes twice the physical drive capacity to support a LUN. RAID 5 can require anywhere from 8% to 33% more physical drive capacity to support a LUN, depending on the stripe size.
|
 |
|
