Modelling Zoned RAID Systems Using Fork-Join Queueing Simulation

Abigail S. Lebrecht, Nicholas J. Dingle, William J. KnottenbeltEPEW 2009 · DOI

TL;DR

This paper introduces a flexible fork-join queueing simulation model for zoned RAID systems (RAID 01 and 5). The model accurately predicts I/O request response times by incorporating realistic disk drive behaviors and workload characteristics, validated against real-world measurements. This work provides a valuable tool for system designers to evaluate RAID configurations and performance under various conditions.

What Problem Does This Paper Address?

RAID (Redundant Array of Independent Disks) systems are fundamental to modern storage, offering high reliability and performance. However, their actual performance can vary significantly based on configuration and workload. Accurately modeling RAID performance is crucial for ensuring Quality of Service (QoS) requirements are met, especially in complex environments like virtualized storage. Existing analytical models often struggle to capture the nuances of real-world RAID behavior, and detailed low-level simulations can be computationally expensive and difficult to parameterize. This paper addresses the need for a flexible, yet accurate, simulation model that bridges this gap, allowing for efficient evaluation of RAID systems, particularly those utilizing zoned disk drives.

Methods

The core of this work is a fork-join queueing simulation model. This model extends a single-disk M/G/1 queue simulation, which itself incorporates realistic seek time, rotational latency, and data transfer time distributions, accounting for zoned disk characteristics (where outer tracks have more sectors and are faster to transfer from).

The fork-join structure is key:

  • Fork: An incoming I/O request is split into multiple subtasks, each destined for a different disk in the RAID array.
  • Parallel Service: Each subtask then queues and is serviced by its respective single-disk simulator.
  • Join: Once all subtasks for a given request have completed, they are synchronized at a join point, and the original request is considered finished.

The authors specifically developed extensions for:

  • RAID 01: Handling mirrored stripes, including efficient read operations (half from primary, half from mirror) and double the subtasks for writes.
  • RAID 5: Modeling parity calculations and updates, including read-modify-write operations for partial stripe writes. This involves complex logic for pre-reading data and parity, calculating new parity, and then writing the updated data and parity.
  • Workload Modeling: The simulator supports both constant and variable request sizes, as well as bulk arrivals, reflecting real-world I/O patterns.
  • Rotational Positioning Optimization (RPO): The model incorporates RPO, where disks reorder queued jobs to minimize head movement, especially important under heavy loads. This is achieved by sampling multiple positioning times and choosing the minimum.

Datasets

The paper does not introduce new benchmark datasets. Instead, it relies on:

  • Device Measurements: The authors conducted experiments on a real Infortrend A16F-G2430 RAID system with four Seagate ST3500630NS disks. This real-world data was used to validate the simulation model's accuracy.
  • Synthetic Workloads: The simulation was evaluated using synthetic workloads characterized by Markovian arrivals, constant and geometrically distributed request sizes, and bulk arrivals.

Key Insights and Ideas

  • Fork-Join as a Natural Fit for RAID: The fork-join queueing model elegantly captures the parallel nature of RAID operations, where a single request is broken down and processed across multiple disks simultaneously before rejoining.
  • Zoned Disks Matter: Accounting for zoned disk characteristics (variable sector density and transfer rates across tracks) is crucial for accurate single-disk modeling, which then propagates to RAID performance.
  • RAID 5 Write Complexity: Simulating RAID 5 writes, especially partial stripe writes, is significantly more complex than reads due to the need for parity calculations and read-modify-write operations. The paper details a sophisticated approach to model this.
  • RPO is Critical Under Load: Rotational Positioning Optimization has a substantial impact on performance, particularly at high arrival rates and long queue lengths. Ignoring RPO leads to significant underestimation of response times.
  • Model Validation is Key: The extensive validation against real hardware measurements demonstrates the model's practical utility and accuracy across various RAID levels and workload types.

Key Results

The paper presents numerous cumulative distribution function (cdf) plots comparing simulation results against real-world measurements for RAID 01 and RAID 5.

  • RAID 01: Excellent agreement is observed between simulation and measurements for both Markovian and bursty arrivals, with constant and geometrically distributed request sizes.
  • RAID 5: Good agreement is generally seen for read operations. For write operations, the simulation tends to underestimate measurements, which the authors attribute to potential unmodeled RAID controller overheads (like parity computation). However, the model captures qualitative trends well, especially for larger partial stripe writes.
  • RPO Impact: Figure 7 clearly shows that enabling RPO in the simulation significantly improves accuracy at higher arrival rates (0.06 requests/ms), demonstrating its importance.

Limitations and Future Work

Limitations: * RAID Controller Overheads: The authors acknowledge that RAID controller overheads, particularly for parity computation in RAID 5, are not fully factored into the simulation and may contribute to the underestimation of response times for write operations. * Caching: The simulation does not currently model disk or RAID system caching mechanisms. * Sequential Access: The current model primarily focuses on random accesses.

Future Work: * Other RAID Levels: Extending the simulator to model other RAID levels. * Mixed Workloads: Incorporating streams of both read and write requests, as well as sequential accesses. * Controller Overheads: Investigating and incorporating more detailed RAID controller overheads. * Caching: Adding caching mechanisms to the simulation model. * Analytical Model Validation: Using the simulator to validate and improve existing analytical queueing models for RAID systems.

Summary updated

BibTeX
@inproceedings{lebrecht2009modelling,
  author = {Abigail S. Lebrecht and Nicholas J. Dingle and William J. Knottenbelt},
  title = {{Modelling Zoned RAID Systems Using Fork-Join Queueing Simulation}},
  booktitle = {EPEW},
  year = {2009},
  doi = {10.1007/978-3-642-02924-0\_2},
}
Figure