White Papers
Dell HPC NFS Storage Solution High Availability Configurations with Large Capacities
31
InfiniBand random read performance Figure 14.
6.4. Metadata tests
From past experience, it was expected that metadata test results would be very similar for 10GbE and
IPoIB; however that was not observed in this environment. For both networks the general curve of the
graphs remained the same, but the absolute performance varied depending on the number of
concurrent clients in the tests. For some client counts IPoIB performed better than 10GbE, but for
others, 10GbE performed better than IPoIB, and the variation was observed to be statistically
significant.
This release of the solution includes updated firmware on the storage arrays, larger capacity
configurations, a newer version of the operating system and XFS file system. With this release of the
solution, metadata performance is much higher than in the previous version showing that both IPoIB
and 10GbE are able to scale. However, for the cases where 10GbE performance is better than IPoIB, it
is possible that the solution is now stressing the IPoIB implementation more than in previous versions
and thus exposing some issues. That being said, the metadata performance of the solution is impressive
and is expected to meet requirements of most clusters. Results for 10GbE will be presented in a blog
post on www.hpcatdell.com at a later date. The InfiniBand results are presented and discussed in this
section.
Figure 15, Figure 16 and Figure 17 show the results of file create, stat and remove operations,
respectively. Recall that the HPC compute cluster has 64 compute nodes. In the graphs below each
client executed a maximum of one thread for client counts up to 64. For client counts of 128, 256 and
512, each client executed 2, 3 or 4 simultaneous operations.
Similar performance for the 144TB and 288TB configurations indicates that the number of disk spindles
is not a factor. The results demonstrate the scalability of the NSS-HA solution for such operations. As
0
2000
4000
6000
8000
10000
12000
1 2 4 8 16 32 48 64
IOPS
Number of concurrent clients
IB Random Read Performance
144 TB -- NFSv3 288 TB -- NFSv3