Vault 2017

The Open-Channel SSD kernel subsystem is maturing, the first block-device FTL (pblk) is ready for upstream kernel, and the liblightnvm user-space library is ready for use. This talk will cover an introduction to Open-Channel SSDs, an update on current progress, and what the community is working on. We show how pblk is initialized, and show its performance on an Open-Channel SSD. At last, we provide an overview of the LightNVM architecture, and how one can develop for Open-Channel SSDs.

Flash based storage supports the concept of discarding data in blocks without actually overwriting, a concept that helps with the internal wear level and data placement algorithms. Linux has supported this concept, which has different names in different storage protocols (trim, unmap, deallocate), for a long time. But the concept of "online" or live discards that notify the device instantly after the deletion of data in the file system has only seen limited traction in Linux, mostly due to the severe performance degradation caused by it.

This talks explains optimizations to the file system and block layer to allow better batching and asynchronous execution of discard requests, as well how the file system block allocator can better be aware of ongoing discards.

It will also explore how discarding of data overlaps with fast zeroing operations, and why it really shouldn't at the interface level.

Jun He has performed a detailed vertical analysis of application performance atop a range of modern file systems and SSD FTLs. He has formalized the "unwritten contract" that clients of SSDs must follow to obtain high performance, and conduct the analysis to uncover application and file system designs that violate the contract. The analysis, which utilizes a highly detailed SSD simulation underneath traces taken from real workloads and file systems, provides insight into how to better construct applications, file systems, and FTLs to realize robust and sustainable performance. In this presentation, Jun He will present the methodology of the vertical analysis, as well as the insights learned from applications (e.g., LevedlDB, RockDB, SQLite-WAL, SQLite-RollBack, and Varmail), file systems (e.g. ext4, F2FS, and XFS) and SSD simulations.

I/O completion polling is a very effective technique for reducing I/O latency under certain workloads. Recent changes to the Linux kernel have greatly improved the efficiency of the I/O polling code in two major areas: (1) lower code overhead for small raw block device direct I/O accesses and (2) CPU load cost of I/O polling through "hybrid" polling.
This presentation will explain in detail these changes and show many performance evaluation results using fast NVMe SSDs to illustrate how effective polling can be compared to a regular interrupt based I/O completion scheme. The gains obtained with recent optimizations will also be numerically detailed.

NVM Express has brought PCIe attached storage to the mainstream, and we're seeing this fast interface replacing solid state storage on slower, albeit more mature, buses. There is a strong user expectation on maintaining feature parity with those mature interfaces, and surprise hot plug is the one feature that trips up testing most often. This is not surprising as many software and hardware components that were previously not in the storage hot plug path need to come together to harmoniously work in any given situation. This presentation will explain the most recent software and hardware enhancements to handle corner case and error conditions, as well as remaining gaps and limitations.