Key Technologies of Enterprise Flash Storage Systems: Global Load Balancing, Multi‑Stream Partitioning, and End‑to‑End I/O Prioritization

As a flash storage system, I/O priority, global load balancing, and multi‑stream partitioning are core technologies; this article continues from previous posts and, based on a mainstream Active‑Active architecture, introduces the implementation forms of an all‑flash storage system.

In traditional ALUA SAN storage, each LUN is bound to a specific controller, requiring careful planning to achieve load balancing, but varying workload pressures make balanced distribution difficult.

Major commercial storage products (Dell EMC, HDS, HPE, Dorado) adopt a Symmetric Active‑Active design, using balancing algorithms and a global cache so that LUNs have no fixed ownership and each controller processes its own I/O requests directly.

1. Global Load Balancing – Host I/O requests are assigned to a controller by hashing the LBA; multipath software, front‑end shared interface modules, and controllers negotiate the same hash algorithm to distribute requests intelligently. Without Huawei multipath or shared interface modules, the controller still hashes the LBA to forward the request to the appropriate controller, ensuring balanced load.

Multipath software distributes host read/write requests to the front‑end interface module, which then forwards them to the selected controller.

For write requests, the global cache stores data locally and mirrors it to a partner node; for read requests, the cache is checked first—if hit, data is returned immediately, otherwise data is read from the storage pool, placed in the cache, and then returned to the host.

2. Multi‑Stream Partition Technology – SSDs use NAND flash composed of chips, blocks, and pages (4 KB or 8 KB). Because NAND must be erased at the block level before writing, moving valid data during erase causes write amplification.

Multi‑stream write classifies data and stores different classes in separate blocks, increasing the likelihood that all pages in a block are either valid or garbage, thereby reducing the amount of data moved during erasure, lowering write amplification, and improving SSD performance and endurance.

Without multi‑stream, hot and cold data are mixed in the same block, leading to larger data movement during garbage collection.

Dorado implements three‑class data segregation (metadata, newly written data, and garbage‑collected valid data) to further reduce write amplification, improve GC efficiency, and extend SSD lifespan.

3. End‑to‑End I/O Priority – Controllers tag I/O requests with priority levels; based on these tags, the CPU scheduler, resource scheduler, and queues enforce end‑to‑end priority guarantees.

With priority control, resources are allocated according to I/O type, preventing contention; high‑priority reads can preempt ongoing writes or erase operations, reducing read latency jitter.

Inside the SSD, erase latency is typically 5–15 ms, write latency 2–4 ms, and read latency tens to hundreds of microseconds; when a flash chip is busy with write or erase, reads must wait, causing significant latency variation.

Reference : OceanStor Dorado flash product introduction.