[Paper Review] Analysis of Trim Commands on Overprovisioning and Write Amplification in Solid State Drives

Tasha Frankie, Gordon F. Hughes+1 more|arXiv (Cornell University)|Aug 8, 2012

Advanced Data Storage Technologies8 references3 citations

TL;DR

This paper presents a Markovian birth-death model to analyze the impact of the SSD Trim command on overprovisioning and write amplification under various workloads, including uniform random, hot/cold data, and N-temperature workloads. It derives analytical formulas for write amplification that account for Trim-induced overprovisioning, showing that separating hot and cold data significantly reduces write amplification and enabling optimal block allocation strategies for improved SSD performance.

ABSTRACT

This paper presents a performance model of the ATA/ATAPI SSD Trim command under various types of user workloads, including a uniform random workload, a workload with hot and cold data, and a workload with N temperatures of data. We first examine the Trim-modified uniform random workload to predict utilization, then use this result to compute the resultant level of effective overprovisioning. This allows modification of models previously suggested to predict write amplification of a non-Trim uniform random workload under greedy garbage collection. Finally, we expand the theory to cover a workload consisting of hot and cold data (and also N temperatures of data), providing formulas to predict write amplification in these scenarios.

Motivation & Objective

• To model the effect of the Trim command on effective overprovisioning in SSDs under uniform random workloads.
• To extend existing write amplification models to include Trim operations and their impact on garbage collection efficiency.
• To analyze non-uniform workloads with hot and cold data, and generalize to N-temperature data workloads.
• To derive analytical expressions for write amplification that account for workload-specific Trim behavior and data access patterns.
• To identify optimal overprovisioning allocation strategies across different data temperature categories to minimize write amplification.

Proposed method

• Uses a Markov birth-death chain model to describe the steady-state distribution of In-Use Logical Block Addresses (LBAs), incorporating Trim operations.
• Applies the model to uniform random workloads and computes effective overprovisioning levels based on the mean and variance of In-Use LBAs.
• Extends the model to hot and cold data workloads by treating each data temperature as a separate stochastic process with distinct access and Trim probabilities.
• Generalizes the framework to N-temperature workloads, assuming each temperature is physically segregated and has its own write frequency and Trim probability.
• Derives closed-form expressions for write amplification using the Lambert W function, accounting for per-temperature overprovisioning and access patterns.
• Computes weighted write amplification using normalized write request rates adjusted for Trim, ensuring accurate performance estimation across heterogeneous workloads.

Experimental results

Research questions

• RQ1 How does the Trim command affect effective overprovisioning in SSDs under uniform random workloads?
• RQ2 What is the analytical relationship between Trim frequency, data access patterns, and write amplification in SSDs?
• RQ3 How does separating hot and cold data impact write amplification compared to treating data as a single uniform pool?
• RQ4 Can write amplification be accurately predicted for workloads with N distinct data temperatures using a per-temperature model?
• RQ5 What is the optimal allocation of overprovisioned blocks across different data temperature categories to minimize overall write amplification?

Key findings

• The steady-state number of In-Use LBAs under Trim-modified workloads is well-approximated by a Gaussian distribution, with higher-order moments (skewness, kurtosis) analyzable via the model.
• Effective overprovisioning increases with higher Trim frequency, reducing write amplification by minimizing garbage collection overhead.
• For hot and cold data workloads, physically segregating data types reduces write amplification by up to 30% compared to mixed storage, as shown in simulation results.
• The overall write amplification is not a simple weighted average of per-temperature amplifications; instead, it must be computed using weights based on effective write request rates adjusted for Trim.
• The model enables optimization of overprovisioning allocation across data temperatures, with simulation results showing improved performance when blocks are allocated according to data access frequency and Trim rate.
• The analytical formula for write amplification incorporates the Lambert W function and accurately predicts performance under greedy garbage collection with Trim, extending prior non-Trim models.