Data Sheet Auto-Tiered Storage Conquering I/O Bottlenecks with Hybrid Storage

Overview For more than a decade, storage system When data is requested from ZFS, it performance has remained rather stagnant first looks to the ARC; if it is there, it can while drive capacities and application be retrieved extremely fast (typically in performance demands steadily have nanoseconds) and provided back to the increased. The result of this trend is application. The contents of the ARC are an expensive problem: Storage users balanced between the most recently used are forced into buying expensive hard (MRU) and most frequently used (MFU) disk drives (HDDs) to get a moderate data. performance boost (by reducing I/O Level-Two ARC (L2ARC): The L2ARC latency) and/or forced into over-buying lives in SSDs. In concept, it is an extension capacity in order to meet performance of the ARC. Without an L2ARC, data that requirements. could not fit in the ARC would have to be With the advent and decreasing price of retrieved from HDDs when requested. That flash, storage vendors are integrating is when drive speed makes a difference, it into their products to solve this but the performance difference between problem. ZFS technology is leading the “fast” (e.g., 15k-RPM) and “slow” (e.g., industry in its ability to automatically and 7,200-RPM) is in terms of latencies intelligently use flash in a storage system measured as a few milliseconds or that offers the appropriate capacity and several milliseconds; both are dramatically performance capabilities at a total cost slower than ARC accesses measured in that is dramatically lower than most legacy nanoseconds. L2ARC, in SSDs, fits nicely storage systems. between the two—both in terms of price and performance. Buying hundreds of gigabytes of flash is The ZFS Hybrid Storage Pool cheaper than the same capacity of DRAM ZFS is a robust, scalable with (though still more expensive today than Hybrid Storage Pools features not available in other file systems HDDs), and flash’s I/O latencies typically available today. One of these revolutionary are measured in microseconds—slower features is the ZFS Hybrid Storage Pool than DRAM but still far faster than even (HSP), which allows you to combine “high-performance” HDDs. The L2ARC is DRAM, SSDs, and spinning HDDs into an populated by data first placed in the ARC accelerated storage medium. Below we will as it becomes apparent that the data might explore each of these offerings. get squeezed out of the ARC, and not every piece of data that existed in ARC will Adaptive Replacement Cache (ARC): make it to the L2ARC (those that do not The ARC lives in DRAM. It is the first would be retrieved from HDDs instead, if destination for all data written to a ZFS requested); the algorithms that manage pool, and it is the fastest (lowest-latency) L2ARC population are automatic and source for data read from a ZFS pool. intelligent. Data Sheet

ZFS Intent Log (ZIL): The ZIL is used Hard Disk Drives (HDD): With the to handle synchronous writes—write ARC, L2ARC, and ZIL/SLOG providing operations that are required by protocol the bulk of the performance from a ZFS (e.g., NFS, SMB/CIFS) to be stored in a Hybrid Storage Pool, spinning drives non-volatile location on the storage device are relegated to the job they do well— before they can be acknowledged—to providing lower-performance, higher- ensure data stability. Databases usually density, low-cost storage capacity. Until the require transactions to be on stable day that flash competes with HDDs on a storage devices before continuing, so dollar-per-gigabyte front, the right balance reducing the latency of synchronous writes of DRAM and flash for performance, and has a direct impact on performance. HDDs for capacity, results in a total cost of ownership (TCO) that is less—both initially ZFS can do this by using placing the ZIL and over the long-term—than solving both on a SSD. All writes (whether synchronous or asynchronous) are written into the requirements using all flash or all HDDs. ARC in DRAM, and synchronous writes A New Storage Parameter: are also written to the ZIL before being Working Set Size acknowledged. For legacy storage systems, sizing means Under normal conditions, ZFS regularly determining necessary capacity, IOPS, bundles up all of the recent writes in the and throughput and then performing some ARC and flushes them to the spinning simple math to determine the number of drives—at which point the data in the ZIL is spindles that could provide those numbers. no longer relevant (because it now exists Working Set Size on its long-term, non-volatile destination) As the industry moves towards more and can be replaced by new writes. The sophisticated caching methodologies in ZIL only is read from when synchronous storage systems, a new parameter for writes in the ARC are unable to be written expressing storage needs has become to spinning disk—like after a power failure evident: The Working Set Size (WSS) can be described as the subset of total data or controller failover—at which point ZFS Conclusion reads the ZIL and places that data onto that is actively worked upon (e.g., 500GB the spinning drives as intended. One of this quarter’s sales data out of a total ZFS hybrid storage pools intelligently might compare this concept to non-volatile database of 20TB). combine DRAM, flash, and hard disk drives to achieve the right balance of RAM (NVRAM) from storage vendors like Knowing the WSS makes it possible to cost and performance for any given NetApp, but where NVRAM uses batteries size ARC, L2ARC, and even HDDs more working set, while reducing the need that can wear out and have other issues, accurately, but few applications today have for administrators to constantly monitor write-optimized SLC (single-level cell) flash an awareness of WSS. devices do not need batteries. storage for I/O bottlenecks. And while NVRAM scalability is limited By reducing both read and write latency to available slots, adding SLOGs is as with the use of flash in a ZFS hybrid easy as adding HDDs. Like L2ARC, the storage pool, we end up with a system that ZIL/SLOG is managed automatically and performs far better than legacy storage intelligently by ZFS: Writes that need it are systems, while having a much lower total accelerated, without any additional effort cost of ownership (TCO). by the administrator.

Nexenta Systems is the leading supplier of enterprise-class , Inc. Nexenta Systems EMEA OpenStorage solutions. Its flagship -only platform, 455 El Camino Real Camerastraat 8, NexentaStor, delivers high-performance, ultra-scalable, cloud- Santa Clara, CA 95050 1322 BC, Almere and -optimized storage solutions. www.nexenta.com The Netherlands www.facebook.com/nexenta Copyright © 2012 Nexenta® Systems, twitter.com/nexenta ALL RIGHTS RESERVED Rev. 080712