HARMONIC Privacy Engine

Docker

A simple docker-compose build --build-arg v=□ --build-arg n=□ --build-arg m=□ (with v being the MP-SPDZ version, n the maximum size of the matrices and m the maximum size of the populations) followed by docker-compose up (maybe including --force-recreate) should be enough
The Swagger UI will then be at 127.0.0.1/docs

There is a very basic test/benchmark script (test.py)

Benchmarks

Ubuntu 22.04.3 LTS
2 x Intel Core2 DUO E8400 @ 3GHz

SlotMachine: MPyC 0.8
HARMONIC: MP-SPDZ 0.3.7

Results in seconds for test.py @ 78a6bd35
matrix size 100; population size 500
variable one-way delay (tc qdisc add dev eth0 root netem delay □ms)

Clear (fitness value of a single configuration)

latency	SlotMachine	HARMONIC Shamir	HARMONIC Replicated Field
0 ms	0.63	0.55	0.54
1 ms	-	0.58	0.66
2 ms	-	0.69	0.66
3 ms	-	0.75	0.73

Order (sorted fitness values of 500 configurations)

latency	SlotMachine	HARMONIC Shamir	HARMONIC Replicated Field
0 ms	100.15	10.84	1.93
1 ms	-	19.15	4.68
2 ms	-	32.07	7.97
3 ms	-	41.78	11.03

Classification (500 configurations, 2 classes)

latency	SlotMachine	HARMONIC Shamir*	HARMONIC Replicated Field*
0 ms	11.22	0.86	1.21
1 ms	-	1.52	4.28
2 ms	-	2.26	7.43
3 ms	-	3.00	10.66

^* without "improvement over last run", "best run so far"

lap_solver (exact solution; average of 10 runs with random matrices)

latency	SlotMachine	HARMONIC Shamir	HARMONIC Replicated Field
0 ms	-	-	175
1 ms	-	-	650
2 ms	-	-	-
3 ms	-	-	-

Benchmarks on l3dss2302

Ubuntu 24.04.3 LTS
16 x Intel(R) Xeon(R) W-2245 CPU @ 3.90GHz
MP-SPDZ 0.4.1, rep_field
bench.py @ 13cfd282
bench_latency.py @ e7736f7c
bench_exacty.py @ 05ca03e0

Order

Population	20x20	40x40	60x60	80x80	100x100
100	0.44s	0.46s	0.47s	0.45s	0.50s
200	0.60s	0.61s	0.63s	0.64s	0.63s
300	0.74s	0.75s	0.77s	0.77s	0.78s
400	0.84s	0.87s	0.89s	0.88s	0.90s
500	0.99s	1.01s	1.04s	1.04s	1.06s

Classification (75%)

Population	20x20	40x40	60x60	80x80	100x100
100	0.54s	0.56s	0.56s	0.58s	0.57s
200	0.63s	0.70s	0.63s	0.64s	0.66s
300	0.68s	0.69s	0.70s	0.72s	0.73s
400	0.76s	0.80s	0.79s	0.80s	0.80s
500	0.82s	0.83s	0.87s	0.87s	0.90s

Buckets (5)

Population	20x20	40x40	60x60	80x80	100x100
100	1.16s	1.20s	1.17s	1.13s	1.20s
200	1.80s	1.86s	1.84s	1.82s	1.84s
300	2.46s	2.47s	2.46s	2.45s	2.48s
400	3.08s	3.04s	3.08s	3.09s	3.10s
500	3.69s	3.73s	3.73s	3.77s	3.79s

Quantiles (5)

Population	20x20	40x40	60x60	80x80	100x100
100	0.69s	0.70s	0.70s	0.69s	0.71s
200	0.92s	0.95s	0.94s	0.94s	0.98s
300	1.15s	1.18s	1.20s	1.22s	1.22s
400	1.38s	1.39s	1.40s	1.41s	1.43s
500	1.57s	1.61s	1.63s	1.67s	1.68s

Top (10)

Population	20x20	40x40	60x60	80x80	100x100
100	0.46s	0.48s	0.48s	0.49s	0.49s
200	0.59s	0.62s	0.61s	0.63s	0.63s
300	0.73s	0.75s	0.76s	0.77s	0.77s
400	0.87s	0.88s	0.90s	0.90s	0.92s
500	0.98s	1.03s	1.03s	1.04s	1.05s

Latency (100x100 matrix, population 500, 2 repetitions)

Latency is one-way delay from tc qdisc add dev eth0 root netem delay □ms

Method	0	1	2	5	10
order	1.04s	2.40s	3.77s	7.70s	14.19s
classification	0.88s	3.53s	5.86s	12.68s	23.99s
buckets	3.71s	24.02s	42.40s	96.66s	186.30s
quantiles	1.64s	3.92s	6.17s	12.80s	23.74s
top	1.04s	2.38s	3.70s	7.53s	13.93s

Exact solutions (100x100 matrix, 10 repetitions)

Latency is one-way delay from tc qdisc add dev eth0 root netem delay □ms

Latency	lap_solver	munkres
0	52s	> 600s
1	550s	...