White Papers

22 PS Series Asynchronous Replication Best Practices and Sizing Guide | BP1012
6.7 Latency and TCP window size effects
For the purpose of this discussion, latency is how long it takes a packet of data to travel from one point to
another across the network. Latency is inherent in any network, including an iSCSI-based SAN. Typically,
iSCSI SAN latencies are quite small, and usually measured in microseconds or milliseconds. Several factors
affect latency in a storage system, including the time it takes to retrieve data off a hard disk, cache hit ratios,
and the speed of the connection to a host system. In our tests, we wanted to understand how the latency
caused by a WAN link might affect asynchronous replication.
Distance will add latency — the round trip (send and acknowledgement) for a data packet exchange will take
longer as distance separates the devices. The maximum possible speed that data can travel is equivalent to
the speed of light in a vacuum: 186,282 miles per second. In practice, data speed is attenuated by the cable
medium. In the case of fiber optic cables, the glass fiber slows the light signal down to approximately 124,000
miles per second or about 199,560 km/second. Table 3 shows the approximate distance a data packet can
travel in a given time across fiber optic cables (values in the table should be doubled for round-trip
calculations):
Latency induced by distance
Time
Approximate distance traveled through fiber optic cables
Milliseconds
Miles
Kilometers
1
124
200
10
1,240
1,996
20
2,480
4,007
50
6,200
9,978
100
12,400
19,956
Of course, there are other factors that affect a network packet’s round-trip time. For example, the amount and
type of network switching and routing devices it encounters along the way will all contribute incrementally to
the overall link latency. In many cases, the device latencies will add more to overall link latency than distance
alone.
In the lab test environment, a WAN emulator was used to inject various amounts of latency into the replication
network links. A volume was then replicated across the link, and the time it took to complete the replication
was measured. These tests used a 10 GB volume and it was ensured that the entire volume was overwritten
each time (100 percent of the data had changed) before creating a new replica.