(A) memory management layer for disaggregated data center

Abstract

The popularity of memory-intensive applications has increased the demand for various systems running them. Existing solutions, such as swapping and memory over-provisioning, fail to support the memory-intensive application. Disk swapping significantly reduces performance of the memory-intensive application due to too frequent evictions. On the other hand, memory over-provisioning incurs low memory utilization which results in high memory cost. Disaggregated data center is proposed to mitigate these limitations. It uses a fast and efficient inter-connection such as RDMA to allocate resources from the remote pool on demand. Prior work suggests disaggregated data center libraries both in userspace and in form of kernel block device driver. Userspace libraries incur additional cost for integrations due to library-specific APIs, whereas kernel block device drivers are not able to exploit user application's information. In this work, We propose a userspace memory management library that mitigates these limitations. It enables each process to use remote memory as same as a local one. Unlike similar resource disaggregation libraries, it offers a POSIX-like API for easy integration. In addition, since the library shares its address space with the user application, it is possible to exchange information between the library and application easily. This introduces additional optimization opportunities. To verify the effectiveness of the proposed library, we implemented its prototype. For basic performance comparisons, we conduct a microbenchmark. It shows that our implementation brings $5 - 6\times$ better performance compared to Linux's disk swapping. Finally, we integrate Memcached, which is a well-known memory-intensive application, with our solution and analyze its performance to evaluate the ease of integration and performance benefits when it comes to the real-world application. The integration only takes only 10 lines of code modification, while it presents $1.16 - 2\times$ higher throughput (TPS) than Linux's disk one.