With each page we are progressing into more difficult scenarios. The ‘overlapped regions’ is one of the most complex access patterns. It is much more demanding of the caching algorithm. What we are simulating is a scenario that would be found typically in a large dataset on a typical server. The areas of the disk that are hot data are going to be mixed in with other data that is less frequently accessed. For ease of understanding, we will call these overlapped regions cold-cool-warm-hot. This really stresses the RAID controller as it has to sift through and decipher what information is needed the most and which is less desirable. Here the data that needs to be cached isn’t cut and dry and that is where the intelligent caching comes into play. Not all data needs to be cached so selective caching will be the best approach.
For this testing, we have used identical settings to the other tests that we have run.
This approach uses 4 individual threads accessing four different regions of the disk itself.
Thread 1- 650 GB range with QD 1
Thread 2- 600 GB range with QD2
Thread 3- 550 GB range with QD4
Thread 4- 20 GB range with QD 128
The fourth thread was intentionally designed to be able to fit into the cache of the controller and, with its higher QD, it becomes the most accessed data and the data that we are monitoring for acceleration. As you can see, this is a much more intensive benchmark that is designed to test very high usage patterns from multiple threads.
The results are as expected with intelligent caching sorting out which data is more important. This performance speaks volumes to the benefits to be had from write caching and also to the optimizations in the algorithm that have been made. Here the performance advantages of using multiple SSDs in the caching volume is also readily apparent.
A quick click will bring the image up to full size.