Go to main content

Prior measurements of ARM TrustZone latency focus on bare-metal monitors or hy-pervisor abstractions rather than the full overhead of a production Trusted ExecutionEnvironment (TEE). This thesis benchmarks Open Portable TEE (OP-TEE) on twoAMD FPGA-SoC platforms, the PYNQ-Z2 (Zynq-7000, Cortex-A9, ARMv7, 667MHz) and the AUP-ZU3 (Zynq UltraScale+, Cortex-A53, ARMv8, 1.2 GHz), mea-suring world-switch latency and end-to-end secure peripheral access through a TrustedApplication (TA). The delay was decomposed into the portion spent switching worldsand the portion spent on AXI transactions. The AUP-ZU3 outperforms the PYNQ-Z2 by 3.0x on SMC round-trip latency and 4.2x on PTA AXI read cost. Expressedin clock cycles derived from wall-clock time, the AUP-ZU3 completes a world-switchround trip in approximately 63,200 cycles compared to 105,000 on the PYNQ-Z2.Normalizing the 3.0x SMC speedup by the 1.8x clock-frequency ratio leaves a further1.66x not explained by clock frequency. This residual is hypothesized to come froma cleaner Exception Level model in ARMv8 than the modes of ARMv7, as well ashighly optimized firmware routing. On both platforms, OP-TEE dispatch overheadaccounts for the majority of secure peripheral access latency. On the PYNQ-Z2, theAXI bus transaction itself accounts for less than 15% of total PTA AXI read cost,while it accounts for over 30% of total PTA AXI read cost on the AUP-ZU3. Forengineers migrating between device generations, the world-switch path is the primarylatency concern, and optimization concerning minimization of OP-TEE world-switchinvocations is the most effective way to increase trusted application performance onARM FPGA-SoCs.

Metric
From
To
Interval
Export
Download Full History