Hardware Reference
In-Depth Information
10000
$WORK
$SCRATCH
1000
100
10
1
FIGURE 7.6: File system usage growth after starting production operations
in January 2013. The SCRATCH system started from being nearly empty to
amassing 1 PB of data in approximately two months.
performance and Figure 7.5(b) presents a time history of daily I/O measure-
ments during August 2013 (after the purge process was initiated). In these
results, there was a fairly constant response, albeit lower than the measured
values when the file system was relatively empty. The average aggregate per-
formance over the 52 measurements during this period was 71 GB/s and this
value provides a realistic expectation of daily performance using 512 clients
in production operations.
7.3 Conclusion
Like many HPC sites across the world, parallel I/O performance and sta-
bility remains a key component in the overall design of large-scale systems
at TACC. Indeed, significant burn-in time and performance evaluation is de-
voted to the I/O subsystem during the deployment phase. Fortunately, much
of the low-level disk and RAID testing can be performed in parallel with the
compute system and interconnect installation and validation.
While peak system IO rates remain useful for overall system characteri-
zation, it is important for users to understand that these peak numbers may
not be indicative of the level of performance they can expect to achieve in