Optimizing SSD Write Performance: Insights from the NoWA Pattern

Modern Solid State Drives (SSDs) have fundamentally changed how we store data, but they often hide a complex internal mechanism that can lead to performance degradation and hardware wear. For database engineers and systems programmers, understanding the gap between how software writes data and how the Flash Translation Layer (FTL) mesma manages it is critical for maximizing throughput and longevity.

Recent research presented at VLDB, titled "How to Write to SSDs," explores this gap, introducing a method to virtually eliminate write amplification. This is particularly relevant for large-scale databases like MySQL and PostgreSQL, where the mismatch between page sizes and physical flash blocks leads to significant overhead.

The Problem: Write Amplification (WAF)

Write amplification occurs when the amount of data physically written to the flash memory is significantly greater than the amount of data the host software actually requested. This happens because flash memory cannot be overwritten in place; it must be erased in large blocks before new data can be written to pages within those blocks.

When a database updates a small piece of data (e.g., an 8KB or 16KB page), the SSD's internal controller must perform a read-modify-write cycle. This process moves existing valid data to a new location and erases the old block, creating a massive amount of internal data movement. As the drive fills up, this "garbage collection" process becomes more aggressive, leading to a performance cliff where write speeds plummet and the drive wears out faster.

The NoWA Pattern: Achieving WAF = 1

The core contribution of the research is the introduction of the NoWA (No Write Amplification) pattern. The most impressive aspect of this approach is that it guarantees a Write Amplification Factor (WAF) of 1, even when the device is at full utilization.

Unlike traditional write patterns, NoWA aligns software writes with the physical characteristics of the flash memory. By ensuring that the SSD controller doesn't have to perform internal data movement to reclaim space, the system can maintain peak performance regardless of how much data is stored on the drive.

Real-World Database Implications

The research analyzed the impact of these write patterns on common databases, including MySQL and PostgreSQL. The findings suggest that these patterns can lead to the development of more optimized storage table formats.

However, the implementation of these patterns varies across different database engines:

• MySQL: Uses a default page size of 16KB.
• PostgreSQL: Uses a default page size of 8KB.
• LeanStore: Uses a page size of 4KB.

These differences in page size significantly affect how the SSD handles writes. For instance, some critics noted that PostgreSQL's approach to writing pages—writing to the Write-Ahead Log (WAL) first and then to data files—creates a sequential write pattern for the WAL, which is generally more SSD-friendly than random writes to the data files.

Industry Perspectives and Alternatives

The discussion surrounding this research highlights several existing industry solutions and alternative approaches to the same problem:

Enterprise Storage Solutions

Enterprise-grade storage systems often mitigate write amplification by using large buffers of NVRAM (8GB or more). These buffers consolidate writes and flush them to the SSDs in large, sequential chunks, effectively acting as a a software-defined NoWA layer.

Zoned Namespaces (ZNS) and Zoned XFS

Some experts pointed out that the same goals of data placement are being pursued at the file system layer through Zoned XFS. This is a Linux file system designed specifically for Zoned SSDs, which allow the host to manage data placement more explicitly, reducing the internal overhead of the FTL.

SMR Hard Drives

Interestingly, the parallel can be drawn to Shingled Magnetic Recording (SMR) hard drives. SMR drives require large sequential writes to avoid the performance penalties associated with overlapping tracks. Optimizing for SMR drives often mirrors the optimization strategies used to optimize for SSDs.

Conclusion

Achieving a WAF of 1 is a significant milestone for database performance. While the research provides a blueprint for a new way to write to SSDs on commodity hardware, the future of database storage engines may lie in a hybrid approach: combining the explicit data placement of Zoned XFS with the design principles of the NoWA pattern to create storage systems that truly understand the physical reality of their hardware.