Characterizing the Performance and Scalability of Many-core Applications on Virtualized Platforms

Clouds have become attractive to applications, because of its low cost and on-demand computing model with the use of virtualization technologies. With the continual increasing number of cores per chip, it should be an emergence to study and improve the scalability of virtualized platforms. This paper tries to make a study on the horizontal scalability 1 of a set of parallel applications on virtualized platforms. By executing and profiling such software on a virtual machine configured with different number of cores on a commodity multi-core machine with 48 cores, we find several performance bottlenecks inside the Xen virtual machine monitor under different paging modes (e.g., direct paging mode and nested paging mode). After a detailed profiling and analysis, we propose several remedies with only less than 100 LOCs to avoid most of the bottlenecks, which result in a performance improvement ranging from 1.1X to 9.42X for a virtual machine configured with 32 cores 2 . The performance scalability is also notably improved. One speculative conclusion from this study is that, though there might be some scalability issues within current virtual machine monitors, some of them should be relatively easy to be refined for commodity multicore platforms.