We analyzed some storage workloads of the FUJITSU K5 cloud service, which was built using the
OpenStack platform, to clarify how to handle these workloads effectively. We retrieved these storage workloads and
analyzed them from the viwe point of both temporal and spatial access locality. For temporal access locality, we
classified these workloads as two input/output (IO) access patterns. One was that the number of IO accesses was
high only for a specific time and the other was that the number of IO accesses was almost stable. For both temporal
and spatial access locality, we found that almost all workloads included IO concentrations. Such concentrations
are aggregations of IO accesses and appear in narrow regions of a storage volume and continue for periods of up
to an hour. Therefore, we studied how to apply cache methods by using a cache simulator. The cache hit ratios
varied with workload, but almost half of the workloads had low cache hit ratios because they included few page-level
regularities. Automated tiered storage with fast memory and slow flash storage (ATSMF), which was we proposed
in our previous study, can handle IO concentration effectively because its migration algorithm found the entire area
of IO concentration and migrated it from slow storage to fast storage if the migration improves user’s performance.
We predict that the cache hit ratios of ATSMF will be much higher than the half workloads of the Least Recently
Used and Adaptive Replacement Cache algorithms. Therefore, we should develop a technique that combines ATSMF
and traditional caching. From the results of temporal access locality, we can a share fast storage area among several
workloads when these workloads have different times when the number of IO accesses is high. Moreover, we should
preferentially allocate a fast storage area to IO concentrated workloads, which accounts for only 1.2% of the total
workloads, because these workloads include more than half of all IO accesses.