Reducing DRAM footprint with NVM in Facebook

Abstract

Popular SSD-based key-value stores consume a large amount of DRAM in order to provide high-performance database operations. However, DRAM can be expensive for data center providers, especially given recent global supply shortages that have resulted in increasing DRAM costs. In this work, we design a key-value store, MyNVM, which leverages an NVM block device to reduce DRAM usage, and to reduce the total cost of ownership, while providing comparable latency and queries-per-second (QPS) as MyRocks on a server with a much larger amount of DRAM. Replacing DRAM with NVM introduces several challenges. In particular, NVM has limited read bandwidth, and it wears out quickly under a high write bandwidth. We design novel solutions to these challenges, including using small block sizes with a partitioned index, aligning blocks post-compression to reduce read bandwidth, utilizing dictionary compression, implementing an admission control policy for which objects get cached in NVM to control its durability, as well as replacing interrupts with a hybrid polling mechanism. We implemented MyNVM and measured its performance in Facebook's production environment. Our implementation reduces the size of the DRAM cache from 96 GB to 16 GB, and incurs a negligible impact on latency and queries-per-second compared to MyRocks. Finally, to the best of our knowledge, this is the first study on the usage of NVM devices in a commercial data center environment.

Faster than Flash: An In-Depth Study of System Challenges for Emerging Ultra-Low Latency SSDs

“…This polling method is implemented in Linux 4.4 kernel, but from 4.10 kernel, Linux supports a hybrid polling [23,49]. The hybrid polling sleeps for a while in an attempt to reduce the number of polling the queues.…”

“…The study proposed to employ a polling-based synchronous I/O service in high performance systems. Similarly, [14,23,43,49,54,68,76,79] advocate to employ the polledmode system for fast NVMe devices. To use the polling in an efficient manner, [23,49,79] proposed different types of a hybrid polling, which combines the current interruptdriven method with the polling by being aware of application characteristics or runtime environment.…”

“…Similarly, [14,23,43,49,54,68,76,79] advocate to employ the polledmode system for fast NVMe devices. To use the polling in an efficient manner, [23,49,79] proposed different types of a hybrid polling, which combines the current interruptdriven method with the polling by being aware of application characteristics or runtime environment. While these studies are mostly performed by simulation or DRAM emulation, of which the actual datapath exists on the memory controller side rather than block storage controller side.…”

AutoTM: Automatic Tensor Movement in Heterogeneous Memory Systems using Integer Linear Programming

“…Use of heterogeneous memory to reduce DRAM has been studied in the past. Facebook has used SSDs to reduce the DRAM footprint of databases [13]. Bandana [14] uses SSD based persistent memory to store deep learning embedding tables [10] with DRAM as a small software cache.…”

Persistent Buffer Management with Optimistic Consistency

“…Second, persistent data structures [1,4,9,14,18] enable direct fine-grained writes to NVM by issuing out-of-place writes and relying on instruction ordering (SFENCE) and eagerly flushing cache lines (CLFLUSH/CLWB) to make the operation visible (usually by flipping a validity bit). Third, buffer managers were proposed to integrate NVM with the storage hierarchy below DRAM and above SSD [5,15,16]. These approaches access data on DRAM, and therefore they have full control of when data is persisted.…”