What is offloaded predicate filtering latency modeling?
However, to use it effectively, you must understand the Latency Model. Offloading isn't "free"; it trades off local compute cycles for interconnect and hardware-logic latency.
To decide whether to offload a query, we model the total latency ($L_{total}$) for both paths.
Where $N$ is the number of rows and $L_{CPU\_Instr}$ is the time to retire the filter instruction.
Where $n$ is the number of surviving rows and $L_{HW\_Logic}$ is the silicon-level latency of the FPGA/ASIC filter.
The Win: Offloading wins when the time saved by not transferring rejected rows over the PCIe bus ($L_{PCIe\_Transfer}$) is greater than the hardware handoff overhead.
Before the storage hardware can filter data, the Database must send the "Predicate" (e.g., Salary > 50000) to the NIC or Storage Cell.
The 2026 Reality: In modern systems using iDB over RDMA, this handoff is sub-microsecond. However, if the hardware requires a full firmware interrupt to load a new filter, $L_{Handoff}$ can spike, making offloading inefficient for very small tables.
Hardware filters (like those in Exadata X11M or AMD Pensando DPUs) use specialized "Stream Processors."
Latency Pattern: These are generally O(1) per packet. Unlike a CPU, which might stall on a branch misprediction while checking a row, the hardware filter processes data at "Line Rate." Whether the data matches or not, the latency through the filter remains a constant, deterministic few nanoseconds.
The most dramatic impact on the model is Selectivity (the ratio of $n/N$).
If your query filters out 99% of data, the PCIe bus latency ($L_{PCIe}$) is reduced by 99%.
In a 1M TPS system, this prevents Bus Saturation, allowing other transactions to fly through while the heavy "Scan" is quietly handled by the storage hardware.
| Metric | Local CPU (No Offload) | Hardware Offload |
| Data Volume | Moves Everything to RAM. | Moves Only Results to RAM. |
| CPU Instruction Retirement | High (Busy sifting). | Zero (Idle/Ready). |
| Tail Latency (P99) | Jittery (Cache misses). | Flat (Line-rate hardware). |
| Optimal Use Case | Small lookups / High hits. | Massive scans / Low hits. |
When modeling offloaded predicate latency for your cluster:
Check the Interconnect: Are you on PCIe Gen5/6? Higher bus speeds reduce the relative benefit of offloading unless your data volume is massive.
Evaluate Complexity: Can your hardware filter handle complex AND/OR logic? If the hardware "bails out" and sends the data back to the CPU for complex filtering, you've paid the $L_{Handoff}$ tax for nothing.
Monitor "Bytes Saved": Use hardware telemetry to measure the Data Reduction Ratio. If you aren't reducing data by at least 2x, the latency of the handoff might outweigh the gains.
Offloaded predicate filtering is the "Turbocharger" of the 1M TPS database. By moving the "Search" to the "Data," you transform a high-latency Data Movement problem into a low-latency Parallel Processing problem. In 2026, the most successful systems are those that treat the CPU as a "Manager" and the Storage Silicon as the "Worker."