Both the pagepool size and maxFilesToCache influence the cache usage.
This indicates that a proper value for maxFilesToCache (mFTC) that fits to the
pagepool size is important.
In addition, consider the number of concurrently open files for a workload when sizing
mFTC. Otherwise, the pagepool might not be optimally utilized.
Figure 1 shows some
unfavourable effects when scaling the values for mFTC.Figure 1. XFS versus Spectrum Scale (different pagepool sizes and mFTC values)
Again, the medium-to-high workload with 750 virtual ICN users is used.
The blue bar shows the read I/O rate for XFS (XFS has around 8 GiB Linux™ page cache available).
For Spectrum Scale there are three measurement series done scaling the pagepool sizes and
mFTC values. The pagepool size starts at 1 GiB (default) and is increased up to 8
GiB in 2GiB steps. For each pagepool size three runs with different mFTC values
(4k, 20k and 100k) are done represented by a colored line for each mFTC value.
Two Unfavourable Scenarios That Might Occur
Scenario 1: maxFilesToCache is set too high for a pagepool size.
For example, this is the case for pagepool sizes <= 4 GiB and mFTC 100k (red
line). The pagepool holds too many metadata and cached files for its size to work efficiently. It
becomes ineffective, which results in high read I/O rates (> 100MiB/sec for this example).
However, the larger the pagepool becomes then the better fits a large mFTC
value. For the 8 GiB pagepool in this measurement series mFTC 100k fits better and
better.
Scenario 2: maxFilesToCache is set too small for a pagepool size.
For example, this happens for pagepool sizes > 4 GiB and mFTC 4k (yellow line).
Larger pagepool sizes have no effect any more because mFTC limits the number of files that can be
cached in the pagepool. As a result, the read I/O rate remains at a certain level even when
enlarging the pagepool.
The value of 20k for mFTC fits well for the workload and pagepool size of 8 GiB
(green line). Compared to XFS the read I/O rate is at a similar level for this combination (41 vs.
49 MiB/sec).