CONFIDE : Content-based Fixed-sized I / O Deduplication

Due to increased adoption of virtualization-based systems, there is an increase in inherent content similarity of systems like email servers, virtualized servers and file servers. Inherent content similarity refers to the same content being present at physically different locations. Harnessing content similarity across different physical locations can help avoid duplicate disk I/O requests that fetch the same content repeatedly. This is referred to as I/O deduplication. Generally, caches are used to store most recently or frequently accessed blocks so as to save time of disk accesses. These caches are looked up based on block number, hence referred as block-based caches. Existing work on I/O deduplication maintains an additional content-based cache in order to serve more disk I/O requests from cache than before. It shows that any given cache can be better utilized if used as a content-based cache rather than block-based. However, the extra cache introduces cache exclusivity concerns. In our work, we incorporate intelligent I/O redirection within the storage virtualization engine of the device to manipulate the underlying block-based cache itself as a content-based cache. We build and evaluate CONFIDE, a storage read-access optimization that identifies content similarity across fixed-sized blocks. This work is applicable to any device containing a storage virtualization layer — SAN volume controller, Shared object store, or Hypervisor, and our target implementation for the scope of this paper is a virtualized host. The CONFIDE system maintains a metadata store and performs read I/O request redirection such that the underlying cache behaves like a contentbased cache. Our trace-based evaluation reveals that CONFIDE always performs equal to or better than Vanilla system, achieving upto 20% better cache-hit ratios and 80% higher number of disk reads averted.