Databases Reference
In-Depth Information
Storage Latency
350
300
250
200
150
100
50
0
1
2
4
8
16 32
Outstanding I/O
64
128
256
FIGURE 4-11
NOTE
The goal of less than 10 milliseconds of latency is no accident. SQL
Server best practices call for no more than 20 milliseconds of delay. If you imple-
ment synchronous data replication with AlwaysOn Availability Groups, you
can't tolerate more than 10 milliseconds of delay; and many applications are
even more sensitive!
The remainder of this chapter explores how to accurately measure performance and latency.
Techniques for establishing performance baselines are demonstrated, and then you will examine
how the application of advanced storage pooling technology is changing how database systems are
designed.
Storage Performance Counters
Windows Performance Monitor (perfmon) allows Windows Server users to capture storage perfor-
mance metrics. For the purposes of storage monitoring, you utilize the LogicalDisk performance
monitor object. Both logical and physical disk counters deliver storage performance metrics. The
logical disk counters show the performance of a specii c partition, while the physical disk coun-
ters cover the entire LUN (a Logical Unit Number is a term that describes a storage volume that is
hosted by a storage controller). Table 4-1 shows a list of the available Windows storage performance
counters.
The Average Disk Sec/Read and Write counters measure the time it takes for an input output (I/O)
operation to be sent from the server to the storage system and back. This latency measure is the
single biggest indicator of I/O system health. Reads and writes are treated separately because most
storage systems perform one operation faster than the other. If you are using a storage array with a
battery-backed cache, it will often write in just a few milliseconds, whereas a random read will
take longer.
Search WWH ::
Custom Search