Cold-Cache Tests; Figure 11. Metadata File Remove Performance - Dell PowerEdge EL Manual

Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
50000
45000
40000
35000
30000
25000
20000
15000
10000
5000
0
1

3.4. Cold-cache tests

For all the test cases discussed in the previous sections, the file system was unmounted and re-
mounted from the I/O clients and the NFS server between test iterations. This was done to eliminate
the impact of caching in the client and server RAM memory, and to present true disk performance.
However, with DFC there is another layer of caching – the DFC cache pool or the SSDs. DFC treats the
SSDs plus the backend virtual disk as part of the DFC configuration. Thus, unmounting the file system
does not necessarily evict the data in the DFC cache. On a subsequent re-mount of the file system, the
last used data is likely to be accessed from the SSD.
A cold-cache is when the data being accessed is not in the SSD cache and has to be retrieved from
backend storage. In all our test cases, all writes are cold-writes since they are fresh writes and not a
re-write of existing data. However all reads are likely to be cache hits since the SSD cache is larger
than the total I/O size. In a production cluster, a worst-case read scenario could arise when a read
request is issued for data that was already flushed out of the SSD cache (say, to make room for newer
requests). The results presented in this section simulate such a cold-cache read.
To simulate a cold-cache, caching was disabled on the backend disk after every write test. This
ensured that the data in the SSD cache was flushed out the backend disk. Additionally the SSDs were
removed from the DFC configuration. Then the SSDs were re-added, and caching re-enabled on the
backend disk and the read test cases executed. Results are presented in Figure 12 and Figure 13. Tests
were conducted with WB mode only. As seen in previous sections, WB and WT mode do not impact read
tests. A cold-cache also does not impact write tests. Consequently, the graphs below present only cold-
cache read results.
Figure 11. Metadata file remove performance
File remove
2 4 8
Number of concurrent clients
baseline
16 32 48 64
DFC-WB DFC-WT
128 256 512
22