Tune-up tips for arrays
				Split a "hot" subset of LUNs across controllers and multiple RAID groups. Some applications are cache friendly (highly sequential) and some aren't (highly random), and can adversely affect each other when run on the same subsystem. If available, use partitioning to isolate applications and increase performance. Make sure there are enough Fibre Channel pipes between the host(s) and the subsystem to support the workload. If you have a high-throughput workload, your host bus adapter transfer size should match or exceed the LUN's segment size. Array performance optimization begins when you're deciding which subsystem configuration options to buy. High-random with low-throughput workloads can get by with today's SATA drives because transfer rates aren't much of a concern. Spreading the LUNs across a number of RAID groups or a larger RAID stripe size (and therefore more spindles) yields more balanced activity against all drive spindles. For highly sequential, high-throughput and high-write workloads, disabling write-back caching (enabling write-through) can improve performance. In write-through mode, write data bypasses the cache and goes straight to the subsystem's disk.

Disk drives come in a number of interfaces (Serial ATA [SATA], Serial-attached SCSI [SAS] or Fibre Channel [FC]), capacities (gigabytes) and speeds (rpm). Storage arrays typically include drives with three or more capacity points (73GB, 146GB and 300GB) and at two or more speeds (10,000 rpm or 15,000 rpm).

High random I/O workloads need high rpm drives. These drives typically offer more I/O per drive spindle than lower rpm drives. High-sequential workloads can probably get by with slower rpm drives because sequential I/O doesn't seek as much. For high-throughput workloads, high rpm drives may offer higher MB/sec transfer rates than lower rpm drives, but this often isn't an issue because even slower drives have enough bandwidth to sustain typical workloads. Faster drives typically cost approximately 50% more than slower drives at similar capacities and interface types, so price may be an additional factor to consider.

Today, FC rules for high throughput and any highly sequential workloads where data transfer is a dominant activity, although some SATA vendors may quibble with this statement considering the new SATA 3Gb/sec transfer rates. High-random with low-throughput workloads can get by with today's SATA drives because transfer rates aren't much of a concern. However, highly random workloads need faster drives that are available only with FC. Some drive vendors have recently introduced fast drives supporting SAS and FC, so it won't be long before fast SAS drives show up in storage subsystems.

Capacity is mostly irrelevant to I/O performance, but the number of spindles isn't. Highly random workloads can take advantage of more spindles because much of the I/O time is drive seek activity), so if you have the option, buy more spindles. For highly sequential workloads, the number of spindles isn't a concern because sequential I/O requires minimal seek activity. In any case, you must ensure that the set of spindles can support your minimum throughput requirements.

Most I/O subsystems offer configuration options from two to 100 or more front-end interfaces. This is mostly an availability and connectivity concern for highly random workloads, but may be a real concern for high-throughput workloads. Those types of workloads will have high MB/sec rates and you may need to overconfigure host paths to sustain the workload.

< PREV PAGE   |   1  |   2  |   3  |   4  |   5  |   6  |   NEXT PAGE  >