White Papers
Table Of Contents
- Executive Summary (updated May 2011)
- 1. Introduction
- 2. Dell NFS Storage Solution Technical Overview
- 3. NFS Storage Solution with High Availability
- 4. Evaluation
- 5. Performance Benchmark Results (updated May 2011)
- 6. Comparison of the NSS Solution Offerings
- 7. Conclusion
- 8. References
- Appendix A: NSS-HA Recipe (updated May 2011)
- A.1. Pre-install preparation
- A.2. Server side hardware set-up
- A.3. Initial software configuration on each PowerEdge R710
- A.4. Performance tuning on the server
- A.5. Storage hardware set-up
- A.6. Storage Configuration
- A.7. NSS HA Cluster setup
- A.8. Quick test of HA set-up
- A.9. Useful commands and references
- A.10. Performance tuning on clients (updated May 2011)
- A.11. Example scripts and configuration files
- Appendix B: Medium to Large Configuration Upgrade
- Appendix C: Benchmarks and Test Tools
Dell HPC NFS Storage Solution - High Availability Configurations
Page 17
2) Virtual disks are created using RAID 6, with 10 data disks and 2 parity disks. This RAID
configuration provides a good balance between capacity and reliability to tolerate multiple disk
failures.
(7)
3) Virtual disks are created with a segment size of 512k to maximize performance
(7)
. This value
should be set based on the expected application I/O profile for the cluster.
4) Cache block size is set to 32k to maximize performance. This value should be set based on the
expected application I/O profile for the cluster.
5) The RAID controllers read and write caches are enabled to maximize performance by allowing
the controller to prefetch data into the caches based on data access patterns.
(8)
6) Write cache mirroring is enabled between the two PowerVault MD3200 RAID controllers to
protect data if there is a controller failure. Cache mirroring between the controllers ensures
that the second controller can complete the writes to disk.
(7)
This design choice does have a
performance penalty. With cache mirroring enabled, writes are 20-25% slower from server
direct to storage when compared to the case where cache mirroring is disabled. Cache
mirroring has little to no impact on read performance.
NFS Server configuration
7) The XFS file system is mounted with the wsync option. By default XFS uses asynchronous
journaling to prevent the journaling subsystem from slowing down operations. This means that
each operation is signaled as complete before the transaction recording the change has been
written to disk. If the HA service on the NFS server fails over before an operation is completed
to disk it can lead to an inconsistent view between the NFS client and the actual data on the
storage. To avoid this problem, the XFS file system needs to run transactions synchronously and
this is enforced by the wsync option. This reduces XFS metadata performance due to the
increase in journal I/O, but if the server crashes and the HA service fails over to the passive
server, the passive server will recover the transactions correctly.
8) The XFS file system is exported using the NFS sync option. This ensures that the client waits for
each write operation to complete at the NFS server, thus, data integrity can be guaranteed as
much as possible in the event of a server failure. The sync option does have a performance
penalty but is an absolute requirement to ensure data integrity.
9) Increased number of concurrent NFS threads (from a default of 8 to 256) on the servers to
maximize performance.
(3)
10) Change in default OS scheduler from cfq to deadline to maximize performance.
(3)
11) Set large MTU (MTU=8192) for 10 Gigabit Ethernet networks to maximize performance.
(3)
4.4. Functionality Tests
To test the failover of the HA cluster service, several component failures were simulated.
1) Server failure
2) Heartbeat link failure
3) Public link failure