This measurement then includes time to:

1. send input to card reader (PC/SC stack)
2. transmit input data (T=1/T=0)
3. dispatch command and select on-card method (process())
4. execute code predeceasing target operation
5. execute target operation
6. execute trailing code after target operation
7. transmit response (dataOut, status word, (T=1/T=0))
8. receive response (PC/SC stack)

Additionally, target operation usually takes only small fraction of measurement with majority taken by operations we like to exclude from measurement (~one ms vs. ~tens of ms). Situation is additionally worsen by possible non-determistic time fluctuations on a host side.

Note: Measurement fluctuations on host side can be mitigated if simpler host architecture is available (e.g., microcontroller-based card reader).

Note: Very precise measurement of elapsed time can be obtained from a power trace, if start and end of selected operation can be identified. Such a measurement requires access to setup with osciloscope and significant time for identification of target operation. We verified selected operation using this method.

Note: Some operations cannot be meaningfully measured without additional operation executed together (e.g., setKey before target operation Cipher.init). We measure both operations together and then substract time for additional operation.

Note: We intentionally did not exclude outlayer measurements as it may contain interesting information regarding non-deterministic behaviour of a card

Measurement procedure:

• start/stop time measured on host
• repetitions on host with same data (5x, NUM_REPEAT_WHOLE_MEASUREMENT, 10x NUM_REPEAT_WHOLE_MEASUREMENT_KEYPAIRGEN)
• default length of data for on-card operation (256B)
• length of variable data for on-card operation (16-512B)
• repetitions of target operation on card (fixed length => 50x NUM_REPEAT_WHOLE_OPERATION, variable length => 5x (to keep overall running time reasonable))
• perftest_measure method (outer):
  • check if not already measured before
  • try perftest_measure - catch exception, try again, then ask for user intervention (physical remove of card, reupload of applet)
• perftest_measure method (innner):
  • for every repetitions of measurement prepare fresh set of objects on card and reset it (prepare_class_XXX, APDU, not measured)
  • Measure processing time without actually calling measured operation (achieved by setting 0 to number of repetitions; testSet.numRepeatWholeOperation set to 0) - repeat 5x NUM_BASELINE_CALIBRATION_RUNS
    • ResetApplet (APDU, not measured)
    • PerfTestCommand
    • => baseline avg time
  • Measure target operation (testSet.numRepeatWholeOperation = 50x fro fixed data or 5x for variable data):
    
    • ResetApplet (APDU, not measured)
    • PerfTestCommand => time
    • measurement time = time - baseline avg time
    • => min, max, avg measurement time
• prepare_class_XXX (on-card)
  • allocate new objects required for testing target operation
  • erase or set RAM/EEPROM arrays into default values
• perftest_class_XXX (on-card)
  • receive apdu data
  • parse incoming settings (num_repeats, data length...)
  • switch (algorithm type)
  • initializations done only once before for(num_repeats)
  • for(num_repeats)
    • target operation
    • (optional) alternate between two different objects to prevent (too quick) use of existing one when card actually decides not to execute operation as the object is still same (e.g., Cipher.init())
  • end, send apdu out