How AI Manages Mixed Surveillance Workloads
Eun-Kyung Hong, HDD Product Manager at Toshiba Electronics Europe GmbH discusses AI optimisation in mixed surveillance workloads. AI-enabled surveillance has changed the operating model of surveillance storage. Continuous recording remains the foundation of every surveillance system, but AI analysis and metadata processing now compete for the same drive resources. Toshiba developed OptiFrame™ AI for this operating environment, classifying storage requests before determining how cache resources and scheduling should be applied. Every activity within an AI surveillance system generates storage traffic. Continuous recording writes new video to the drive. AI analysis retrieves previously captured footage, and metadata is updated as events are identified and indexed. These activities often occur simultaneously, but they access the same hard disk drive (HDD) in different ways. Video recording depends on long, predictable sequential writes that maintain a steady flow of incoming data. By contrast, By contrast, AI analysis often generates bursts of random reads while accessing previously recorded footage. Metadata updates introduced further small read and write operations. Each access pattern suits its own task, but together they place different demands on the mechanics and cache architecture of the HDD. Why sequential and random access behave differently An HDD delivers its highest sustained write performance when data is written sequentially. Because each block follows the last, the recording heads move smoothly across the media with very little seek activity. This makes sequential writes well suited to the uninterrupted stream of incoming video produced by surveillance cameras. Random reads interrupt that pattern. Instead of reading consecutive blocks, the recording heads move repeatedly to different locations across the media before reading can begin. Each additional seek interrupts the continuous recording pattern and increases the time required to complete the request. Neither access pattern is inherently better than the other. Storage requests carry different priorities A request to retrieve previously recorded footage may arrive while new video is still being written, but delaying the two operations has different consequences. A brief delay while retrieving historical footage may have little effect on an analytics task. Interrupting the recording stream, however, risks losing video that cannot be recreated. The drive’s firmware therefore has to distinguish between requests that can wait and those that require immediate service. The drive must identify each request A continuous recording stream, a request to analyse previously captured footage and an update to surveillance metadata all access the same storage media, yet each has different operational requirements. Treating them identically risks allowing one activity to interfere with another. HDD firmware has to do more than manage data movement. It must identify the type of request and decide how the drive should respond, balancing sequential writes, random reads and metadata activity without interrupting continuous recording. Toshiba’s OptiFrame™ AI firmware performs this classification before applying cache management and command scheduling (Figure 1). Figure 1: OptiFrame™ AI classifies storage requests before applying workload-specific cache management and command scheduling Applying the right response Once a request has been classified, cache behaviour can be adapted to the workload (Figure 2). Continuous recording benefits from keeping incoming video flowing steadily to the media. Analytical requests behave differently, repeatedly retrieving small sections of previously recorded data. OptiFrame™ AI protects the recording stream without preventing analytical processing and metadata operations. Figure 2: OptiFrame™ AI allocates cache resources according to workload type, protecting continuous recording while adapting cache behaviour for AI analysis and metadata operations The scheduler determines the order in which requests are serviced. Recording activity has strict timing requirements because missed video cannot be recovered, whereas many analytical operations can tolerate a short delay without affecting the overall result. The scheduler therefore gives priority to recording when workloads compete for drive resources. Evaluating HDD performance in AI surveillance Continuous recording remains the primary function of a surveillance HDD. Toshiba’s OptiFrame™ AI firmware was developed to support the additional storage demands created by AI analysis and metadata processing, helping the S300 AI surveillance HDD sustain continuous recording AI changes how surveillance storage is evaluated, making workload coordination a key aspect of HDD performance alongside storage throughput.
