 | ||
| Striping for I/O-request-intensive applications | ||
|---|---|---|
| Prev | Best practices for I/O performance tuning | Next |
A good compromise stripe unit size for I/O-request-intensive applications is one that results in a 3% to 5% probability of splitting in a uniform distribution of requests. For example, a 2 KB (four-block) database page size would have an ideal stripe unit size of 100 blocks. This would typically be rounded up to the nearest power of two (128 blocks, or 65,536 bytes) for simplicity.
I/O-request-intensive applications are typically characterized by small (for example, 2 to 16 KB) data transfers for each request. These applications are I/O bound because they make so many I/O requests, not because they transfer large amounts of data.
For example, an application that makes 1,000 I/O requests per second with an average request size of 2 KB uses at most 2 MB per second of data transfer bandwidth. Because each I/O request occupies a disk completely for the duration of its execution, the way to maximize I/O throughput for I/O-request-intensive applications is to maximize the number of disks that can be executing requests concurrently. Clearly, the largest number of concurrent I/O requests that can be executed on a volume is the number of disks that contribute to the volume's storage. Each application I/O request that "splits" across two stripe units occupies two disks for the duration of its execution, reducing the number of requests that can be executed concurrently and thus the efficiency of I/O response.
Therefore, try to minimize the probability that I/O requests "split" across stripe units in I/O-request-intensive applications.
The following factors influence whether an I/O request with a random starting address will split across two stripe units:
The request starting address relative to the starting address of the storage allocation unit (the file extent)
The size of the request relative to the stripe unit size
Most database management systems will allocate pages in alignment with the blocks in a file, so that requests for the first page will almost never split across stripe units. However, database requests for two or more consecutive pages may split across stripe units. In this case, larger stripe unit sizes reduce the probability of split I/O requests. However, the primary objective of striping data across a volume is to cause I/O requests to be spread across the volume's disks. Too large a stripe unit size is likely to reduce this spreading effect.