Hardware Reference
In-Depth Information
TABLE 3.3: IOR-write performance on Blue Waters file systems.
File system
Performance
/home
104.3 GB/s
/project
100.0 GB/s
/scratch
980.7 GB/s
FIGURE 3.3: The total I/O bandwidth, tested at Blue Waters acceptance and
then periodically after, is greater than 1.18 TB/s. [Image courtesy of Celso
Mendes and Michelle Butler.]
ten independent IOR executions. As expected, the /scratch file system, which
is ten times bigger than the other file systems, has nearly ten times more de-
livered bandwidth than /home and /project. Figure 3.3 shows the same data
in graph form.
Another remarkable aspect of the Blue Waters file system is its scalable
architecture. This feature ensures that the file system both maintains its per-
formance characteristics as more clients access it, and delivers increasing per-
formance as more servers are configured. The first property was verified by
running IOR tests on an increasing number of clients; each client was running
on a different node, writing a file with a size of 1 GB to the /scratch file
system. Table 3.4 presents the results of these executions, and shows that the
delivered file system performance does not vary significantly, regardless of the
number of clients in use.
To confirm the file system scalability with the number of servers, Figure 3.4
displays the results of IOR executions, for both \read" and \write" operations,
when the number of accessed file system racks (i.e., servers) in the /scratch file
system is varied. Each IOR job had an number of clients proportional to the
number of servers. As the figure shows, observed performance is nearly linear
