Dell PowerVault MD1200 - Manual (Page 17)

Improving NFS Performance on HPC Clusters with Dell Fluid Cache for DAS
17
• DFC in Write-Back mode (DFC-WB) – This configuration builds on the baseline by adding DFC as
described in Section 2.2, and DFC is configured to operate in Write-Back (WB) mode. WB mode
allows the caching of writes on the cache pool. WB mode requires the data to be written to a
minimum of two PCIe SSDs. Both re-reads and writes are accelerated.
• DFC in Write-Through mode (DFC-WT) – This configuration builds on the baseline, but here DFC
is configured in Write-Through (WT) mode. WT mode forces writes to both the cache and
virtual back end disk simultaneously.
The following sections present the results of the different I/O patterns.
3.1. Sequential writes and reads
The results of the IPoIB sequential write tests are shown in Figure 5. The figure shows the aggregate
throughput that can be achieved when a number of clients are simultaneously writing to the storage
over the InfiniBand fabric.
The results show that baseline configuration can reach a peak write throughout of ~2,000 MiB/s. Recall
that this is with the “sync” NFS export option, and this peak throughput demonstrates how well the
configuration is tuned
4 . With DFC in WB mode, labeled DFC-WB in Figure 5, the throughput measured
was ~400 MiB/s with a peak of ~600 MiB/s. The lower sequential write performance with DFC is due to
two factors: the pure sequential write performance of the SSDs is lower than the storage array and the
write-back cache has to replicate dirty blocks which means every write to the cache has to be written
to two SSDs. Subsequent re-write operations were found to have ~25% higher throughput as the files
are already in the DFC cache.
Switching to WT mode and eliminating the replica blocks doubles the sequential write performance.
Peak performance approaches ~1,000 MiB/s as seen in Figure 5, labeled DFC-WT. Recall that all writes
go directly to the backend disk in addition to being cached on the SSDs in WT mode.
If the I/O pattern is such that there is a large amount of sequential data written to the backend
storage initially followed by subsequent reads, re-reads, and small writes, one method to take
advantage of DFC performance on random workload (Section 3.2) while minimizing the sequential write
performance penalty is to disable caching on the backend disk during the write operation. Although this
might not be an option for production clusters, smaller single-user/single-application environments
might be able to adopt this approach. DFC provides very simple utilities to accomplish this. Once the
data is written to the backend store, for example by a gene sequencer, caching on the backend disk
can be re-enabled. Subsequent reads, re-reads, and writes can benefit greatly from DFC technology.