Efficient Persistence of Financial Transactions in NVM-based Cloud Data Centers

Abstract

Performance and reliability are two core challenges for today's cloud data centers. Emerging non-volatile memory (NVM) technologies, which promise large capacity, high-speed, byte-addressable and persistent memory, can offer mitigating benefits. In particular, business-critical applications in ecommerce, finance, and banking could persist transactions in NVM either as traditional storage or directly as durable memory, after additional development to adapt the applications to the new interface. However, both approaches have very diverse software overheads that in the literature are still scantly compared head-to-head or clearly quantified. In order to shed some light on these issues, we developed a suite of throughput, latency, and scalability tests that focus on the challenge of persisting financial transactions in the form of small and critical parcels of data, a representative challenge for financial cloud data centers. By carrying out benchmarks on a real NVDIMM server, we compare and contrast in detail the performance of the programming framework Mnemosyne with the NVM storage solutions PMFS (a persistent memory file system) and PMBD (a persistent memory block-device). In turn, these are compared with both directly-addressable volatile RAM and a fast NVM Express flash drive (NVMe) as performance baselines. We found that persisting financial transactions with Mnemosyne achieves up to two orders of magnitude better throughput than persisting them in the NVMe, while incurring a performance penalty of 25 percent over volatile RAM. Furthermore, committing transactions in NVM as persistent memory or flat files is up to two orders of magnitude faster than persisting them in databases saved in NVM. Finally, the throughput of writing financial transactions using Mnemosyne is four times higher than PMFS and one order of magnitude higher than PMBD.