HPDA: A hybrid parity-based disk array for enhanced performance and reliability

Bo Mao, Hong Jiang, Suzhen Wu, Lei Tian, Dan Feng, Jianxi Chen, Lingfang ZengACM Trans. Storage 2012 · DOI

TL;DR

This paper introduces HPDA, a novel hybrid disk array architecture that combines SSDs and HDDs to overcome the limitations of traditional RAID configurations for modern storage systems. By intelligently using SSDs for data and HDDs for parity and a write buffer, HPDA significantly boosts performance and reliability, particularly for demanding workloads.

What Problem Does This Paper Address?

Modern computing demands increasingly high performance and reliability from storage systems. While Solid State Drives (SSDs) offer significant speed advantages over traditional Hard Disk Drives (HDDs), they come with their own challenges: poor random write performance, the "erase-before-write" issue, and limited write endurance (flash wear-out). Applying standard RAID schemes to SSDs directly can exacerbate these issues, leading to performance bottlenecks and reduced reliability. Conversely, relying solely on HDDs limits overall performance. This paper addresses the need for a storage solution that leverages the strengths of both SSDs and HDDs to achieve superior performance and reliability, especially for applications with mixed read/write and bursty access patterns.

Methods

The core innovation is the Hybrid Parity-based Disk Array (HPDA) architecture. HPDA combines a group of SSDs (as data disks) and a portion of an HDD (as a parity disk) to form a RAID4-like structure. Crucially, it introduces a RAID1-based mirroring write buffer using a second HDD and the remaining space of the parity HDD. This buffer serves two primary purposes:

  1. Absorbing Small Writes: It temporarily stores small, random write requests, which are a bottleneck for SSDs. Writes to this buffer are sequential and fast, avoiding the erase-before-write penalty.
  2. Surrogate During Recovery: It acts as a temporary data store during disk failure recovery, ensuring data availability.

The system operates with three key modules:

  • Monitor: Identifies I/O patterns, distinguishing between sequential and random writes.
  • Director: Routes I/O requests to either the RAID4 array (SSDs) or the mirroring buffer (HDDs).
  • Reclaimer: Periodically moves data from the mirroring buffer back to the SSDs during idle periods, optimizing storage utilization and performance.

This design aims to mitigate SSD write limitations and leverage HDD's capacity and cost-effectiveness for parity and buffering.

Datasets

The paper utilizes two types of datasets for evaluation:

  1. IOmeter: A synthetic benchmark tool used to test various RAID architectures under different access patterns (random/sequential reads/writes) and request sizes.
  2. Financial Traces (OLTP Application I/O): Real-world workload traces obtained from the Storage Performance Council, representing typical online transaction processing (OLTP) applications. These traces are used for performance evaluation in both normal operation and during disk failure recovery.

The paper notes that these traces exhibit burstiness and idleness, which are key characteristics HPDA is designed to exploit.

Key Insights and Ideas

  • Write Buffer is Key: The mirroring write buffer is the linchpin of HPDA, effectively decoupling fast sequential writes from slow random writes and the erase-before-write penalty inherent in SSDs.
  • Leveraging HDD Strengths: Instead of trying to force HDDs into roles where they are inherently slow (like high-performance data storage), HPDA uses them strategically for parity and as a large, fast write buffer, where their capacity and sequential read/write capabilities are advantageous.
  • Workload Awareness: HPDA's design explicitly considers the bursty and idle nature of real-world workloads, using idle periods to reclaim data, thereby improving overall system responsiveness.
  • Reliability Boost: By offloading parity calculations for small writes to HDDs and using mirroring for a write buffer, HPDA avoids the accelerated wear-out on SSDs that can occur with traditional parity-based RAID schemes, leading to higher Mean Time To Data Loss (MTTDL).
  • Cost-Performance Sweet Spot: HPDA offers a compelling balance, achieving performance close to SSD-only solutions while maintaining better reliability and potentially lower cost than pure SSD RAID configurations.

Key Results

  • Performance: HPDA significantly outperforms both HDD-based RAID5 and SSD-based RAID5 in various scenarios.
  • Random Writes: HPDA shows dramatic improvements, outperforming HDD-based RAID5 by up to 512% and SSD-based RAID5 by up to 4558% in IOmeter tests.
  • Random Reads: HPDA performs well, though SSD-based RAID5 can sometimes edge it out due to higher parallelism.
  • Sequential Writes: HPDA performs better than SSD-based RAID5 but can be slightly slower than HDD-based RAID5.
  • Benchmark Traces: HPDA consistently shows the best average response time and reconstruction time during recovery, outperforming HDD-based RAID5 by up to 22.3x and SSD-based RAID5 by up to 8.4x.
  • Reliability: Reliability analysis shows HPDA has a higher Mean Time To Data Loss (MTTDL) compared to both HDD-based and SSD-based RAID arrays, especially as disk repair times increase.
  • Performance/Cost & Energy: HPDA offers a good performance/cost ratio, particularly for small requests, and demonstrates better energy efficiency than HDD-based RAID5.

Limitations and Future Work

  • Complexity: The HPDA architecture introduces more complexity than a simple RAID setup due to the multiple components and management modules (Monitor, Director, Reclaimer).
  • HDD Bottleneck for Sequential Writes: While HPDA excels in random writes, its sequential write performance can be limited by the HDD's capabilities, especially when compared to HDD-based RAID5.
  • Future Work: The authors suggest that HPDA's architecture is orthogonal and complementary to other storage system optimizations and can be embedded within them. Further exploration of its integration with other techniques is implied. The paper also highlights the need for continued analysis and evaluation of SSD-based disk arrays.

Summary updated

BibTeX
@article{mao2012hpda,
  author = {Bo Mao and Hong Jiang and Suzhen Wu and Lei Tian and Dan Feng and Jianxi Chen and Lingfang Zeng},
  title = {{HPDA: A hybrid parity-based disk array for enhanced performance and reliability}},
  journal = {ACM Trans. Storage},
  year = {2012},
  doi = {10.1145/2093139.2093143},
}
Figure