Quantum 6-01376-05 User Manual

  StorNext File System Tuning

The Underlying Storage System

StorNext File System Tuning Guide

3

metadata operations throughput. This is easily observed in the hourly 

File System Manager (FSM) statistics reports in the 

 file. For 

example, here is a message line from the cvlog file:

PIO HiPriWr SUMMARY SnmsMetaDisk0 sysavg/350 sysmin/333 sysmax/367

This statistics message reports average, minimum, and maximum write 

latency (in microseconds) for the reporting period. If the observed 

average latency exceeds 500 microseconds, peak metadata operation 

throughput will be degraded. For example, create operations may be 

around 2000 per second when metadata disk latency is below 500 

microseconds. However, if metadata disk latency is around 5 

milliseconds, create operations per second may be degraded to 200 or 

worse.
Another typical write caching approach is a “write-through.” This 

approach involves synchronous writes to the physical disk before 

returning a successful reply for the I/O operation. The write-through 

approach exhibits much worse latency than write-back caching; therefore, 

small I/O performance (such as metadata operations) is severely 

impacted. It is important to determine which write caching approach is 

employed, because the performance observed will differ greatly for small 

write I/O operations.
In some cases, large write I/O operations can also benefit from caching. 

However, some SNFS customers observe maximum large I/O 

throughput by disabling caching. While this may be beneficial for special 

large I/O scenarios, it severely degrades small I/O performance; 

therefore, it is suboptimal for general-purpose file system performance.

0

RAID read-ahead caching is a very effective way to improve sequential 

read performance for both small (buffered) and large (DMA) I/O 

operations. When this setting is utilized, the RAID controller pre-fetches 

disk blocks for sequential read operations. Therefore, subsequent 

application read operations benefit from cache speed throughput, which 

is faster than the physical disk throughput.
This is particularly important for concurrent file streams and mixed I/O 

streams, because read-ahead significantly reduces disk head movement 

that otherwise severely impacts performance. 
While read-ahead caching improves sequential read performance, it does 

not help random performance. Furthermore, some SNFS customers 

actually observe maximum large sequential read throughput by disabling 

caching. While disabling read-ahead is beneficial in these unusual cases,