Improving Application Performance Through System Call Composition

Long-running server applications can easily execute millions of common data-intensive system calls each day, incurring large data copy overheads. We introduce a new framework, Compound System Calls (Cosy), to enhance the performance of such applications. Cosy provides a mechanism to safely execute data-intensive code segments in the kernel. Cosy encodes a C code segment containing system calls in a compound structure. The kernel executes this aggregate compound directly, thus avoiding data copies between user-space and kernelspace. With the help of a Cosy-GCC compiler, regular C code can use Cosy facilities with minimal changes. Cosy-GCC automatically identifies and encodes zerocopy opportunities across system calls. To ensure safety in the kernel, we use a combination of static and dynamic checks, and we also exploit kernel preemption and hardware features such as x86 segmentation. We implemented the system on Linux and instrumented a few data-intensive applications such as those with database access patterns. Our benchmarks show performance improvements of 20–80% for non-I/O bound applications.