Prateek Sharma

As part of my Master's thesis, I've looked at improving the efficiency of inter-VM page deduplication. Page deduplication in virtual environments is really useful to reduce the overall memory footprint of the virtual machines and provides memory-overcommitment. Unfortunately, finding duplicate pages is resource intensive. As part of our efforts, we focussed on improving KSM (Kernel SamePage Merging), the linux/KVM implementation of page deduplication. By exploiting spatial locality among duplicate pages and implementation tweaks, we have managed to reduce the overhead of KSM by more than a factor of two. On AMD processors, we achieve a 10x speedup, thanks to NPT notifications for dirtied pages.

Part of the work on page deduplication is present in the Singleton paper and my thesis. A qemu-patch for deduplication instrumentation is here.

To combat the double caching which happens during disk IO in most virtual machine setups (blocks are resident in both the host and guest page caches), a black-box solution for exclusive caching (called ``Singleton'') has been developed. Singleton uses inter-VM page deduplication to remove redundant pages from the host page cache — the result being smaller host caches, faster disk I/O, and more VMs.

The Singleton paper was accepted for publication at the High Performance Parallel and Distributed Computing(HPDC) conference. (Slides)

The page cache in most operating system remains largely an opaque and mysterious component of the kernel. The per-file page cache is a new page cache architecture which splits the page cache on a per-file basis and gives each file a private, isolated cache. Each file's caching policy (cache size, eviction algorithm) can be set from userspace. This allows better QoS and systemwide performance. The default giant LRU/2 list in linux has been replaced by multiple CAR (Clock with Adaptive Replacement) lists. This improves scalability by removing contention from the lru-list lock and provides a better eviction algorithm (CAR vs LRU/2).

The Per-file page cache is currently being benchmarked. A Patch is available here, and a Technical report is also available.

In case of virtual machines, disk requests go through two I/O schedulers — one in the VM itself, and the other at the hypervisor. The question to answer is : Is Disk scheduling really passe ? Preliminary benchmarks with KVM and VirtIO indicate that the thumb-rule of "smarter scheduler near the applications, simpler near the disks" might not always hold true.

To identify and understand delays and bottlenecks in the I/O path, we have instrumented the VirtIO disk handler to carefully account for all delays at the guest as well as the hypervisor (QEMU + KVM + host kernel).

I am fascinated by non-invasive system-wide profiling and instrumentation, particularly DTrace. An idea which I have been working on is to unify instrumentation on the host and the virtual machines. The primary motivation is to be able to instrument virtual machines with minimal guest operating system intervention.

In setting up rCloud (IIT Bombay's private cloud for research and academic purposes), we tested a large number of management software, and ended up with Cloudstack.