Quantum 6-01376-05 User Manual

Page of 32
  StorNext File System Tuning
The Metadata Controller System
StorNext File System Tuning Guide
11
InodeCacheSize
0
This setting
 
consumes about 800-1000 bytes of memory times the number 
specified. Increasing this value can reduce latency of any metadata 
operation by performing a hot cache access to inode information instead 
of an I/O to get inode info from disk, about 100 to 1000 times faster. It is 
especially important to increase this setting if metadata I/O latency is 
high, (for example, more than 2ms average latency). You should try to 
size this according to the sum number of working set files for all clients. 
The recommended range is 16K to 64K. 
Example: 
InodeCacheSize          16K     # 800-1000 bytes each, default 8K
ThreadPoolSize
0
This setting consumes 512 KB memory times the number specified. 
Increasing this value can improve concurrency of metadata operations. 
For example, if many client processes are executing concurrently, the 
thread pool can become exhausted by I/O wait time. Increasing the 
thread pool size permits hot cache operations to be processed that would 
otherwise be backed up behind the I/O-bound operations. There are 
various O/S limits to the number of threads that can cause fatal problems 
for the FSM daemon, so it's not a good idea to set this setting too high. A 
range from 32 to 128 is recommended, depending on the amount of 
available memory. It is recommended to size it according to the 
max 
threads
 FSM hourly statistic reported in the 
cvlog
 file.
Example:
 ThreadPoolSize          32      # default 16, 512 KB memory per thread
ForcestripeAlignment
0
This setting should always be set to Yes. This is critical if the largest 
StripeBreadth
 defined is greater than 1MB. Note that this setting is not 
adjustable after initial file system creation.
Example: 
ForcestripeAlignment    Yes 
FsBlockSize
0
The optimal settings are in the range of 4K, 8K, 16K, or 32K. Settings 
greater than 32K can adversely impact performance because metadata I/
O operations are performed less efficiently. Any value greater than 4K