Summary. The HP AutoRAID is a hierarchical storage system which caches write-active data using RAID-1 and stores write-inactive data using RAID-5. The main idea is that RAID-1 has higher throughput and lower latency for random writes but requires more storage than RAID-5. The AutoRAID system essentially introduces a new level in the memory hierarchy by using one RAID level (i.e. RAID-1) to cache another (i.e. RAID-5).
The AutoRAID system implements the following features and functionality:
An HP AutoRAID devices comprises the typical RAID hardware: microprocessors, disks, volatile memory, non-volatile memory, parity modules, etc. The disks are divided into (typically 1MB) physical extents (PEX) which are grouped into physical extent groups (PEG). Moreover, each PEX is divided into segments: the atomic units used for mirroring or striping.
The logical data model is based on 64 KB relocation blocks (RB) that are the atomic unit of logical data movement. Each AutoRAID device contains multiple levels of maps (i.e. virtual device table, PEG table, PEX table) to map logical addresses to physical addresses.
The HP AutoRAID device caches reads and caches writes into NVRAM before issuing them to the backend (which allows for things like write coalescing). If data is being written and is in the RAID-1 cache, it is written there. Otherwise, it is promoted from the RAID-5 storage and brought into the RAID-1 cache. Reads and writes issued to the RAID-1 cache are trivial. Reads from the RAID-5 storage is also simple. Writes to the RAID-5 storage are always appended and also batched to reduce the number of times parity bits that have to be read, computed, and written.
Background operations are triggered when the disk is idle and include:
Commentary. - There are a lot of RAID levels and each one has different capacity, performance, and reliability. Moreover, performance depends heavily on workload. This paper presents the AutoRAID system without much explanation for when or why it's expected to perform better than alternatives. For example, is AutoRAID only better for random-write workloads? The microbenchmarks show that it performs well under sequential writes as well, but isn't RAID-1 supposed to perform poorly with sequential writes? - The AutoRAID system is supposed to be easier to configure than traditional RAID devices. The authors argue that typical RAID configuration requires detailed knowledge of the expected workloads which can be difficult to come by. However, the AutoRAID system also depends on expected workload knowledge. The paper admits that the working write set has to be small and slowly changing. Later in the paper, the authors also say "it is fairly easy to predict or detect the environments that have a large write working-set and to avoid them if necessary"; a bit contradictory from earlier. Finally, the summary says "there are workloads that do not suit its algorithms well". - I think the evaluation of the system makes it unclear whether the hierarchical storage or the NVRAM caching is responsible for the performance improvements. - The AutoRAID system logs all I/O operations made to it in order to form disk traces. Implementing this functionality in the disk device is arguably a violation of the end-to-end principle and could have been implemented in the OS. However, they argue that the performance overhead is negligible.