Performance

Performance Considerations

When you use Conclave to build an enclave, Conclave compiles your enclave code into native machine code using GraalVM Native Image. During this process the compiler removes any code that is not used and performs optimisations on the generated codebase.

Entering and exiting an enclave causes a context switch which is fairly expensive in terms of CPU clock cycles. Also, memory accesses inside an enclave can be slower, particularly if memory paging is required. This means that performance inside an enclave will normally be lower than with a regular HotSpot JVM. This table shows the performance difference and how they vary between variety of benchmarks taken from the Computer Language Benchmarks Game.

The configuration used to perform the benchmarks were:

1. Warmup iteration: 10
2. Number of iterations for each benchmark: 100
3. Azure machine series: Standard DC4s v3 (4 vcpus - Intel(R) Xeon(R) Platinum 8370C CPU @ 2.80GHz, 32 GiB memory)
4. Operating system: Ubuntu 20.04.5 LTS
5. Conclave version: 1.3

Note

The "empty" benchmark measures the overhead of entering and exiting the enclave, without doing any work. As entering/exiting triggers various hardware mechanisms designed to block side channel attacks this is expensive relative to the cost of a regular function call. However, this transition cost becomes less relevant when the enclave is doing real work.

Benchmark	SGX	Score (ops/s)	Score Delta (%)	Error (ops/s)	Error (%)
empty	Off	24908588.801	N.A	±42949.048	±0.17
empty	On	15659.463	-99.94	±1247.194	±7.96
binary_trees	Off	1495.911	N.A	±8.387	±0.56
binary_trees	On	675.869	-54.82	±10.270	±1.52
fannkuch	Off	5.103	N.A	±0.011	±0.22
fannkuch	On	4.600	-9.86	±0.003	±0.07
fasta	Off	3.536	N.A	±0.003	±0.08
fasta	On	2.855	-19.26	±0.002	±0.07
himeno	Off	0.394	N.A	±0.002	±0.51
himeno	On	0.266	-3.49%	±0.003	±1.13
mandelbrot	Off	5.045	N.A	±0.007	±0.14
mandelbrot	On	4.523	-10.35	±0.003	±0.07
nbody	Off	1.385	N.A	±0.002	±0.14
nbody	On	1.411	1.88	±0.002	±0.14
pidigits	Off	19.449	N.A	±0.035	±0.18
pidigits	On	14.950	-23.13	±0.050	±0.33
spectral_norm	Off	12.818	N.A	±0.019	±0.15
spectral_norm	On	9.677	-24.50	±0.009	±0.09

Higher scores are better. As you can see, the performance hit overall of using an enclave is highly dependent on what exactly the code is doing (primarily, memory access patterns). The score delta shows the difference in percentage between the same benchmark with SGX turned on and off. As expected, there is a performance hit when SGX is used.