How NVMe FDP Optimizes RocksDB Performance: Validation with DapuStor R6101

RocksDB, a mainstream LSM-Tree key-value storage engine, is widely used in distributed databases, cloud-native storage, and edge computing. However, a long-standing challenge continues to affect system performance and NAND endurance: the host side has visibility into the complete data lifecycle, while traditional block storage interfaces provide no mechanism to communicate this information to the SSD.

As a result, data with different lifecycles can be mixed within the same NAND flash, increasing SSD garbage collection (GC) overhead. NVMe Flexible Data Placement (FDP) provides a standardized mechanism to bridge this semantic gap.

Using the DapuStor R6101 enterprise SSD with native NVMe FDP support, this article explains how FDP works with RocksDB and validates its performance benefits through comparative testing.

Real-World Validation: FDP Optimizes RocksDB Performance

01 | Why RocksDB Data Mixing Leads to Write Amplification

RocksDB generates data with distinct lifecycle characteristics. Write-Ahead Log (WAL) data is short-lived and becomes obsolete after memtables are flushed to SST files. SST files, by contrast, generally remain valid for longer periods, with their lifetimes increasing as they move to lower levels of the LSM tree. Files in upper levels are rewritten more frequently during compaction, while data in lower levels tends to remain valid for longer.

Limited by traditional block storage interfaces, data with different lifecycles is mixed in the same NAND blocks by the SSD's Flash Translation Layer (FTL). When short-lived data expires, long-lived data in the same block must be moved to new blocks for garbage collection.

This repeated valid data movement leads to device-level write amplification and performance jitter. It consumes NAND endurance, SSD bandwidth, and controller resources that could otherwise serve host I/O.

02 | NVMe FDP Data Isolation

NVMe FDP bridges host-side data placement policies with SSD-side NAND allocation.

RocksDB uses Placement Identifiers (PIDs) to map I/O streams with similar lifecycles to corresponding Reclaim Unit Handles (RUHs) in the SSD. WAL is assigned to a dedicated PID, while SST files are assigned to different PIDs based on LSM-Tree levels. The SSD controller directs data associated with each RUH to the corresponding Reclaim Unit (RU).

By isolating data with different lifecycles at the NAND level, FDP effectively reduces data mixing and delivers 3 practical benefits:

 1. Lower Write Amplification: Similar-lifetime data expire at similar times, leaving less valid data to move during garbage collection.

 2. Extended NAND Endurance: Less data movement reduces additional NAND P/E cycles and physical wear in write-intensive LSM-tree workloads.

 3. Improved Performance: Lower garbage collection overhead reduces contention for SSD bandwidth, controller resources, and host I/O queuing, enabling more stable throughput and lower tail latency.

03 | Performance Validation with DapuStor R6101

Test Environment Configuration

Test Workload Configuration

DapuStor R6101 achieves deep data isolation through both hardware and firmware, ensuring stable throughput and effectively contained tail-latency under high-concurrency read/write workloads.

Test Result

1. Unrestricted Write Throughput

When stress test write throughput is unrestricted, FDP increased RocksDB frontend throughput from 163.7 MB/s to 200.3 MB/s (+22.36%). Background compaction read and write throughput also increased by 18.33% and 18.37%, respectively. These results demonstrate that FDP can effectively improve the system's overall capacity for high-concurrency writes.

2. Stress Test Write Throughput at 150 MiB/s

At a constant 150 MiB/s write workload to assess system stability, response times across all percentiles decreased. P99.9 tail latency dropped from 20,890.6 μs to 5,918.3 μs (-71.7%), while stall count fell from 434 million to 25.1 million (-94.22%). Under sustained writes, the system delivers more stable and predictable QoS, reducing the probability of extreme request timeouts.

The test results show that tail latency and system-level stalls in the database engine can be significantly affected by bursty I/O congestion triggered by SSD garbage collection. With native FDP support, the DapuStor R6101 reduces unpredictable NAND-level blocking, allowing more deterministic I/O response to the application layer and reducing stall count and tail latency.

04 | FDP Value Across Diverse Workloads

RocksDB’s I/O model is representative of many LSM-tree-based data-intensive wrokloads. This validation confirms the effectiveness of FDP data isolation and its broad applicability to workloads with diverse data lifecycles.

 1. AI Data-Intensive Workloads: Optimize data placement for training data, caches, and other data with different lifecycles.

 2. Multi-Tenant Cloud Storage: Isolate tenant I/O streams to minimize interference.

 3. CDN Cache Nodes: Separate hot and cold data to reduce garbage collection overhead.

 4. Large-Scale Log and Stream Processing: Provide stable performance under high-frequency write workloads.

The DapuStor R6101 enterprise SSD reduces write amplification, delivering stable performance for high-throughput and low-latency workloads.

Beyond data isolation, FDP also enables tighter host-NAND coordination, allowing SSDs to play a more active role in data lifecycle management.