Quantum 6-01376-05 User Manual

Page of 32
  StorNext File System Tuning
The Metadata Network
StorNext File System Tuning Guide
7
The Metadata Network
As with any client/server protocol, SNFS performance is subject to the 
limitations of the underlying network. Therefore, it is recommended that 
you use a dedicated Metadata Network to avoid contention with other 
network traffic. Either 100BaseT or 1000BaseT is required, but for a 
dedicated Metadata Network there is usually no benefit from using 
1000BaseT over 100BaseT. Neither TCP offload nor are jumbo frames 
required.
It is best practice to have all SNFS clients directly attached to the same 
network switch as the MDC systems. Any routing required for metadata 
traffic will incur additional latency that impacts performance. 
It is critical to ensure that 
speed/duplex
 settings are correct, as this will 
severely impact performance. Most of the time 
auto-detect 
is the correct 
setting. Some managed switches allow setting 
speed/duplex
, such as 
100Mb/full
, which disables 
auto-detect
 and requires the host to be set 
exactly the same. However, performance is severely impacted if the 
settings do not match between switch and host. For example, if the switch 
is set to
 auto-detect
 but the host is set to 
100Mb/full
, you will observe a 
high error rate and extremely poor performance. On Linux the 
mii-diag
 
tool can be very useful to investigate and adjust 
speed/duplex
 settings.
It can be useful to use a tool like 
netperf
 to help verify the Metadata 
Network performance characteristics. For example, if 
netperf -t TCP_RR 
reports less than 15,000 transactions per second capacity, a performance 
penalty may be incurred.
The Metadata Controller System
The CPU and memory power of the MDC System are important 
performance factors, as well as the number of file systems hosted per 
system. In order to ensure fast response time it is necessary to use 
dedicated systems, limit the number of file systems hosted per system 
(maximum 8), and have an adequate CPU and memory.