{ "schema_version": 3, "commit": "dd28f268cc881a68b1155d34f5020c6808164a9d", "run_url": "https://github.com/A3S-Lab/Gateway/actions/runs/31090975521", "generated_at": "2026-08-06T09:57:42Z", "environment": { "runner_image": "ubuntu24-20260720.247.2", "cpu_model": "AMD EPYC 7763 64-Core Processor", "logical_cpus": 4, "memory_mib": 15989, "kernel": "Linux 6.17.0-1020-azure x86_64 GNU/Linux" }, "versions": { "a3s_gateway": "a3s-gateway 1.0.13", "nginx": "nginx version: nginx/1.24.0 (Ubuntu)", "load_generators": { "oha": "oha 1.15.0", "protocol_load": "a3s-gateway dd28f268" } }, "methodology": { "scope": "Same-host comparison across every traffic type supported by the Gateway data plane: HTTP/1.1, HTTPS, HTTP/2, gRPC, SSE, WebSocket, TCP, UDP, OpenAI JSON, and OpenAI streaming.", "trials": 3, "duration_seconds_per_trial": 10, "warmup_seconds": 2, "connections": 64, "http2_concurrency": { "connections": 4, "parallel_streams_per_connection": 16 }, "aggregation": "Median of repeated trials. Product order alternates on each trial for every traffic profile.", "completion_policy": "The duration stops new HTTP work and in-flight HTTP/1.1 requests drain. HTTP/2 and gRPC retain four client connections for the trial; the NGINX fixture raises keepalive_requests to 1,000,000 so its default connection rotation cannot cancel active streams.", "threshold": "Ratios within 3% are marked within threshold; other results state which measured value is higher or which latency is lower." }, "profiles": { "http1-small": { "label": "HTTP/1.1", "traffic": "http", "unit": "requests/s", "load_generator": "oha", "capability_alignment": "equivalent", "workload": "GET, keep-alive, 42-byte JSON response", "proxies": { "a3s-gateway": { "trials": [ { "success_rate": 1.0, "operations_per_second": 38750.361646250894, "average_latency_us": 1649.02694841716, "p50_latency_us": 1549.978, "p90_latency_us": 2521.2479999999996, "p99_latency_us": 3740.229 }, { "success_rate": 1.0, "operations_per_second": 39053.39352193416, "average_latency_us": 1636.1905084363493, "p50_latency_us": 1536.71, "p90_latency_us": 2499.557, "p99_latency_us": 3723.338 }, { "success_rate": 1.0, "operations_per_second": 38359.30408189978, "average_latency_us": 1665.6597540610344, "p50_latency_us": 1567.019, "p90_latency_us": 2541.763, "p99_latency_us": 3812.248 } ], "median": { "success_rate": 1.0, "operations_per_second": 38750.361646250894, "average_latency_us": 1649.02694841716, "p50_latency_us": 1549.978, "p90_latency_us": 2521.2479999999996, "p99_latency_us": 3740.229 } }, "nginx": { "trials": [ { "success_rate": 1.0, "operations_per_second": 52089.64291265093, "average_latency_us": 1226.217796434385, "p50_latency_us": 1147.4479999999999, "p90_latency_us": 1974.038, "p99_latency_us": 3220.202 }, { "success_rate": 1.0, "operations_per_second": 51482.39074350847, "average_latency_us": 1240.771500929106, "p50_latency_us": 1166.453, "p90_latency_us": 2011.157, "p99_latency_us": 3340.4210000000003 }, { "success_rate": 1.0, "operations_per_second": 51389.530609373316, "average_latency_us": 1243.0560159895456, "p50_latency_us": 1172.5749999999998, "p90_latency_us": 2005.2569999999998, "p99_latency_us": 3195.0429999999997 } ], "median": { "success_rate": 1.0, "operations_per_second": 51482.39074350847, "average_latency_us": 1240.771500929106, "p50_latency_us": 1166.453, "p90_latency_us": 2005.2569999999998, "p99_latency_us": 3220.202 } } }, "comparison": { "a3s_to_nginx_throughput_ratio": 0.7526915725283604, "a3s_to_nginx_p50_latency_ratio": 1.3287959309119186, "a3s_to_nginx_p90_latency_ratio": 1.2573191366493173, "a3s_to_nginx_p99_latency_ratio": 1.161488937650495, "positions": { "throughput": "nginx_higher", "p50_latency": "nginx_lower", "p90_latency": "nginx_lower", "p99_latency": "nginx_lower" } } }, "https-http1": { "label": "HTTPS \u00b7 HTTP/1.1", "traffic": "encrypted-http", "unit": "requests/s", "load_generator": "oha", "capability_alignment": "equivalent", "workload": "GET, downstream TLS termination, keep-alive", "proxies": { "a3s-gateway": { "trials": [ { "success_rate": 1.0, "operations_per_second": 35725.82978700081, "average_latency_us": 1788.6167168142092, "p50_latency_us": 1670.387, "p90_latency_us": 2715.986, "p99_latency_us": 4071.626 }, { "success_rate": 1.0, "operations_per_second": 35756.76476785158, "average_latency_us": 1787.1027746712764, "p50_latency_us": 1671.124, "p90_latency_us": 2699.8399999999997, "p99_latency_us": 4002.169 }, { "success_rate": 1.0, "operations_per_second": 35798.54350490669, "average_latency_us": 1784.9479342652094, "p50_latency_us": 1668.412, "p90_latency_us": 2705.802, "p99_latency_us": 4024.939 } ], "median": { "success_rate": 1.0, "operations_per_second": 35756.76476785158, "average_latency_us": 1787.1027746712764, "p50_latency_us": 1670.387, "p90_latency_us": 2705.802, "p99_latency_us": 4024.939 } }, "nginx": { "trials": [ { "success_rate": 1.0, "operations_per_second": 40524.841283804075, "average_latency_us": 1576.5956267109775, "p50_latency_us": 1468.749, "p90_latency_us": 2459.883, "p99_latency_us": 3749.107 }, { "success_rate": 1.0, "operations_per_second": 40200.85049039652, "average_latency_us": 1589.3422288044537, "p50_latency_us": 1460.662, "p90_latency_us": 2575.3999999999996, "p99_latency_us": 4086.3160000000003 }, { "success_rate": 1.0, "operations_per_second": 39839.07274758886, "average_latency_us": 1603.7884590316462, "p50_latency_us": 1480.558, "p90_latency_us": 2528.862, "p99_latency_us": 3929.584 } ], "median": { "success_rate": 1.0, "operations_per_second": 40200.85049039652, "average_latency_us": 1589.3422288044537, "p50_latency_us": 1468.749, "p90_latency_us": 2528.862, "p99_latency_us": 3929.584 } } }, "comparison": { "a3s_to_nginx_throughput_ratio": 0.8894529427031257, "a3s_to_nginx_p50_latency_ratio": 1.1372855402795168, "a3s_to_nginx_p90_latency_ratio": 1.0699682307694134, "a3s_to_nginx_p99_latency_ratio": 1.0242659273856978, "positions": { "throughput": "nginx_higher", "p50_latency": "nginx_lower", "p90_latency": "nginx_lower", "p99_latency": "within_threshold" } } }, "https-http2": { "label": "HTTPS \u00b7 HTTP/2", "traffic": "multiplexed-http", "unit": "requests/s", "load_generator": "oha", "capability_alignment": "equivalent", "workload": "GET, downstream TLS termination, 4 connections \u00d7 16 streams", "proxies": { "a3s-gateway": { "trials": [ { "success_rate": 1.0, "operations_per_second": 39463.87744880787, "average_latency_us": 1613.1043083526502, "p50_latency_us": 1546.0929999999998, "p90_latency_us": 2321.7670000000003, "p99_latency_us": 3216.786 }, { "success_rate": 1.0, "operations_per_second": 38377.98292278018, "average_latency_us": 1658.701628998634, "p50_latency_us": 1593.402, "p90_latency_us": 2383.202, "p99_latency_us": 3263.492 }, { "success_rate": 1.0, "operations_per_second": 39760.679139095024, "average_latency_us": 1600.9961350065985, "p50_latency_us": 1534.739, "p90_latency_us": 2298.663, "p99_latency_us": 3152.0229999999997 } ], "median": { "success_rate": 1.0, "operations_per_second": 39463.87744880787, "average_latency_us": 1613.1043083526502, "p50_latency_us": 1546.0929999999998, "p90_latency_us": 2321.7670000000003, "p99_latency_us": 3216.786 } }, "nginx": { "trials": [ { "success_rate": 1.0, "operations_per_second": 36008.125011347794, "average_latency_us": 1767.831291683389, "p50_latency_us": 1724.795, "p90_latency_us": 2257.979, "p99_latency_us": 3079.351 }, { "success_rate": 1.0, "operations_per_second": 21260.111450929053, "average_latency_us": 2995.6926007080433, "p50_latency_us": 3038.592, "p90_latency_us": 3422.979, "p99_latency_us": 3769.204 }, { "success_rate": 1.0, "operations_per_second": 21927.077653736913, "average_latency_us": 2904.7962598042704, "p50_latency_us": 2826.28, "p90_latency_us": 3350.232, "p99_latency_us": 3735.8410000000003 } ], "median": { "success_rate": 1.0, "operations_per_second": 21927.077653736913, "average_latency_us": 2904.7962598042704, "p50_latency_us": 2826.28, "p90_latency_us": 3350.232, "p99_latency_us": 3735.8410000000003 } } }, "comparison": { "a3s_to_nginx_throughput_ratio": 1.799778250070741, "a3s_to_nginx_p50_latency_ratio": 0.5470416943827221, "a3s_to_nginx_p90_latency_ratio": 0.6930167821213576, "a3s_to_nginx_p99_latency_ratio": 0.8610607357218896, "positions": { "throughput": "a3s_higher", "p50_latency": "a3s_lower", "p90_latency": "a3s_lower", "p99_latency": "a3s_lower" } } }, "grpc-unary": { "label": "gRPC unary", "traffic": "grpc", "unit": "requests/s", "load_generator": "oha", "capability_alignment": "equivalent", "workload": "Empty unary message, downstream HTTP/2 TLS, h2c upstream", "proxies": { "a3s-gateway": { "trials": [ { "success_rate": 1.0, "operations_per_second": 6596.724818742667, "average_latency_us": 9658.306989480632, "p50_latency_us": 2620.514, "p90_latency_us": 42489.902, "p99_latency_us": 44613.886 }, { "success_rate": 1.0, "operations_per_second": 5684.108316160213, "average_latency_us": 11123.271722996677, "p50_latency_us": 2660.96, "p90_latency_us": 42559.965, "p99_latency_us": 80850.369 }, { "success_rate": 1.0, "operations_per_second": 6120.545497546816, "average_latency_us": 10403.759676011488, "p50_latency_us": 2626.393, "p90_latency_us": 42496.026, "p99_latency_us": 82323.329 } ], "median": { "success_rate": 1.0, "operations_per_second": 6120.545497546816, "average_latency_us": 10403.759676011488, "p50_latency_us": 2626.393, "p90_latency_us": 42496.026, "p99_latency_us": 80850.369 } }, "nginx": { "trials": [ { "success_rate": 1.0, "operations_per_second": 2847.4139917758725, "average_latency_us": 22319.20002713581, "p50_latency_us": 40815.416, "p90_latency_us": 42817.291, "p99_latency_us": 45030.747 }, { "success_rate": 1.0, "operations_per_second": 3111.1363502762556, "average_latency_us": 20435.794340332177, "p50_latency_us": 1602.487, "p90_latency_us": 42145.945999999996, "p99_latency_us": 43678.275 }, { "success_rate": 1.0, "operations_per_second": 2848.403739198645, "average_latency_us": 22323.86091686193, "p50_latency_us": 40812.331, "p90_latency_us": 42747.057, "p99_latency_us": 44832.399000000005 } ], "median": { "success_rate": 1.0, "operations_per_second": 2848.403739198645, "average_latency_us": 22319.20002713581, "p50_latency_us": 40812.331, "p90_latency_us": 42747.057, "p99_latency_us": 44832.399000000005 } } }, "comparison": { "a3s_to_nginx_throughput_ratio": 2.1487633277959177, "a3s_to_nginx_p50_latency_ratio": 0.06435292804030233, "a3s_to_nginx_p90_latency_ratio": 0.9941275255510572, "a3s_to_nginx_p99_latency_ratio": 1.8033915383381558, "positions": { "throughput": "a3s_higher", "p50_latency": "a3s_lower", "p90_latency": "within_threshold", "p99_latency": "nginx_lower" } } }, "sse-finite": { "label": "SSE", "traffic": "server-streaming", "unit": "streams/s", "load_generator": "oha", "capability_alignment": "equivalent", "workload": "Finite three-event stream, HTTP/1.1, buffering disabled", "proxies": { "a3s-gateway": { "trials": [ { "success_rate": 1.0, "operations_per_second": 38379.10484257605, "average_latency_us": 1664.8168524846283, "p50_latency_us": 1562.05, "p90_latency_us": 2553.5570000000002, "p99_latency_us": 3855.607 }, { "success_rate": 1.0, "operations_per_second": 38131.453007368866, "average_latency_us": 1675.7627925293373, "p50_latency_us": 1561.31, "p90_latency_us": 2620.0110000000004, "p99_latency_us": 4007.348 }, { "success_rate": 1.0, "operations_per_second": 38816.422598750185, "average_latency_us": 1646.1293750312188, "p50_latency_us": 1547.243, "p90_latency_us": 2513.623, "p99_latency_us": 3808.5060000000003 } ], "median": { "success_rate": 1.0, "operations_per_second": 38379.10484257605, "average_latency_us": 1664.8168524846283, "p50_latency_us": 1561.31, "p90_latency_us": 2553.5570000000002, "p99_latency_us": 3855.607 } }, "nginx": { "trials": [ { "success_rate": 1.0, "operations_per_second": 50804.60000713819, "average_latency_us": 1257.2308619920648, "p50_latency_us": 1185.642, "p90_latency_us": 2026.61, "p99_latency_us": 3246.719 }, { "success_rate": 1.0, "operations_per_second": 49933.24803905409, "average_latency_us": 1279.3224285010472, "p50_latency_us": 1178.3449999999998, "p90_latency_us": 2143.353, "p99_latency_us": 3651.542 }, { "success_rate": 1.0, "operations_per_second": 51242.19799311892, "average_latency_us": 1246.5553853222482, "p50_latency_us": 1165.7620000000002, "p90_latency_us": 2042.3350000000003, "p99_latency_us": 3392.1910000000003 } ], "median": { "success_rate": 1.0, "operations_per_second": 50804.60000713819, "average_latency_us": 1257.2308619920648, "p50_latency_us": 1178.3449999999998, "p90_latency_us": 2042.3350000000003, "p99_latency_us": 3392.1910000000003 } } }, "comparison": { "a3s_to_nginx_throughput_ratio": 0.7554257850112721, "a3s_to_nginx_p50_latency_ratio": 1.325002439862689, "a3s_to_nginx_p90_latency_ratio": 1.2503125099457237, "a3s_to_nginx_p99_latency_ratio": 1.1366125905056643, "positions": { "throughput": "nginx_higher", "p50_latency": "nginx_lower", "p90_latency": "nginx_lower", "p99_latency": "nginx_lower" } } }, "websocket-echo": { "label": "WebSocket", "traffic": "bidirectional-message", "unit": "messages/s", "load_generator": "protocol-load", "capability_alignment": "equivalent", "workload": "64 persistent connections, 32-byte binary echo", "proxies": { "a3s-gateway": { "trials": [ { "success_rate": 1.0, "operations_per_second": 67449.9, "average_latency_us": 948.2637542828085, "p50_latency_us": 912.0, "p90_latency_us": 1324.0, "p99_latency_us": 1735.0 }, { "success_rate": 1.0, "operations_per_second": 67415.6, "average_latency_us": 948.7981639264502, "p50_latency_us": 912.0, "p90_latency_us": 1326.0, "p99_latency_us": 1755.0 }, { "success_rate": 1.0, "operations_per_second": 68286.9, "average_latency_us": 936.6613171779653, "p50_latency_us": 902.0, "p90_latency_us": 1305.0, "p99_latency_us": 1710.0 } ], "median": { "success_rate": 1.0, "operations_per_second": 67449.9, "average_latency_us": 948.2637542828085, "p50_latency_us": 912.0, "p90_latency_us": 1324.0, "p99_latency_us": 1735.0 } }, "nginx": { "trials": [ { "success_rate": 1.0, "operations_per_second": 80410.8, "average_latency_us": 795.3607425868167, "p50_latency_us": 677.0, "p90_latency_us": 1311.0, "p99_latency_us": 3158.0 }, { "success_rate": 1.0, "operations_per_second": 66622.1, "average_latency_us": 960.0640192969, "p50_latency_us": 1109.0, "p90_latency_us": 1530.0, "p99_latency_us": 2078.0 }, { "success_rate": 1.0, "operations_per_second": 71592.3, "average_latency_us": 893.3916203278844, "p50_latency_us": 996.0, "p90_latency_us": 1414.0, "p99_latency_us": 1944.0 } ], "median": { "success_rate": 1.0, "operations_per_second": 71592.3, "average_latency_us": 893.3916203278844, "p50_latency_us": 996.0, "p90_latency_us": 1414.0, "p99_latency_us": 2078.0 } } }, "comparison": { "a3s_to_nginx_throughput_ratio": 0.9421390289179142, "a3s_to_nginx_p50_latency_ratio": 0.9156626506024096, "a3s_to_nginx_p90_latency_ratio": 0.9363507779349364, "a3s_to_nginx_p99_latency_ratio": 0.8349374398460058, "positions": { "throughput": "nginx_higher", "p50_latency": "a3s_lower", "p90_latency": "a3s_lower", "p99_latency": "a3s_lower" } } }, "tcp-echo": { "label": "TCP", "traffic": "layer-4-stream", "unit": "round trips/s", "load_generator": "protocol-load", "capability_alignment": "equivalent", "workload": "64 persistent connections, 32-byte echo", "proxies": { "a3s-gateway": { "trials": [ { "success_rate": 1.0, "operations_per_second": 74968.7, "average_latency_us": 853.0336233654846, "p50_latency_us": 818.0, "p90_latency_us": 1192.0, "p99_latency_us": 1577.0 }, { "success_rate": 1.0, "operations_per_second": 74477.3, "average_latency_us": 858.7518599627, "p50_latency_us": 822.0, "p90_latency_us": 1205.0, "p99_latency_us": 1599.0 }, { "success_rate": 1.0, "operations_per_second": 73824.9, "average_latency_us": 866.3278609249725, "p50_latency_us": 830.0, "p90_latency_us": 1214.0, "p99_latency_us": 1610.0 } ], "median": { "success_rate": 1.0, "operations_per_second": 74477.3, "average_latency_us": 858.7518599627, "p50_latency_us": 822.0, "p90_latency_us": 1205.0, "p99_latency_us": 1599.0 } }, "nginx": { "trials": [ { "success_rate": 1.0, "operations_per_second": 81846.4, "average_latency_us": 781.3343262501467, "p50_latency_us": 609.0, "p90_latency_us": 1335.0, "p99_latency_us": 3299.0 }, { "success_rate": 1.0, "operations_per_second": 80562.7, "average_latency_us": 793.7994866110495, "p50_latency_us": 637.0, "p90_latency_us": 1302.0, "p99_latency_us": 3219.0 }, { "success_rate": 1.0, "operations_per_second": 80682.9, "average_latency_us": 792.57316110353, "p50_latency_us": 698.0, "p90_latency_us": 1186.0, "p99_latency_us": 2342.0 } ], "median": { "success_rate": 1.0, "operations_per_second": 80682.9, "average_latency_us": 792.57316110353, "p50_latency_us": 637.0, "p90_latency_us": 1302.0, "p99_latency_us": 3219.0 } } }, "comparison": { "a3s_to_nginx_throughput_ratio": 0.923086552416931, "a3s_to_nginx_p50_latency_ratio": 1.2904238618524333, "a3s_to_nginx_p90_latency_ratio": 0.9254992319508448, "a3s_to_nginx_p99_latency_ratio": 0.4967381174277726, "positions": { "throughput": "nginx_higher", "p50_latency": "nginx_lower", "p90_latency": "a3s_lower", "p99_latency": "a3s_lower" } } }, "udp-echo": { "label": "UDP", "traffic": "layer-4-datagram", "unit": "round trips/s", "load_generator": "protocol-load", "capability_alignment": "equivalent", "workload": "64 connected sockets, 32-byte datagram echo", "proxies": { "a3s-gateway": { "trials": [ { "success_rate": 1.0, "operations_per_second": 78582.4, "average_latency_us": 813.8396549354563, "p50_latency_us": 708.0, "p90_latency_us": 1238.0, "p99_latency_us": 1787.0 }, { "success_rate": 1.0, "operations_per_second": 77797.7, "average_latency_us": 822.0449023557252, "p50_latency_us": 719.0, "p90_latency_us": 1251.0, "p99_latency_us": 1817.0 }, { "success_rate": 1.0, "operations_per_second": 78360.9, "average_latency_us": 816.1307501572851, "p50_latency_us": 711.0, "p90_latency_us": 1239.0, "p99_latency_us": 1801.0 } ], "median": { "success_rate": 1.0, "operations_per_second": 78360.9, "average_latency_us": 816.1307501572851, "p50_latency_us": 711.0, "p90_latency_us": 1239.0, "p99_latency_us": 1801.0 } }, "nginx": { "trials": [ { "success_rate": 1.0, "operations_per_second": 53011.0, "average_latency_us": 1206.7288638207165, "p50_latency_us": 1155.0, "p90_latency_us": 1837.0, "p99_latency_us": 2712.0 }, { "success_rate": 1.0, "operations_per_second": 52146.3, "average_latency_us": 1226.7434391318272, "p50_latency_us": 1169.0, "p90_latency_us": 1875.0, "p99_latency_us": 2770.0 }, { "success_rate": 1.0, "operations_per_second": 52997.3, "average_latency_us": 1207.051651687916, "p50_latency_us": 1159.0, "p90_latency_us": 1831.0, "p99_latency_us": 2730.0 } ], "median": { "success_rate": 1.0, "operations_per_second": 52997.3, "average_latency_us": 1207.051651687916, "p50_latency_us": 1159.0, "p90_latency_us": 1837.0, "p99_latency_us": 2730.0 } } }, "comparison": { "a3s_to_nginx_throughput_ratio": 1.4785828712028724, "a3s_to_nginx_p50_latency_ratio": 0.6134598792062123, "a3s_to_nginx_p90_latency_ratio": 0.6744692433315188, "a3s_to_nginx_p99_latency_ratio": 0.6597069597069597, "positions": { "throughput": "a3s_higher", "p50_latency": "a3s_lower", "p90_latency": "a3s_lower", "p99_latency": "a3s_lower" } } }, "openai-json": { "label": "OpenAI JSON", "traffic": "ai-request", "unit": "requests/s", "load_generator": "oha", "capability_alignment": "a3s_feature_enabled_vs_nginx_transport", "workload": "Chat Completions JSON validation and forwarding", "proxies": { "a3s-gateway": { "trials": [ { "success_rate": 1.0, "operations_per_second": 36063.928480172945, "average_latency_us": 1771.8386388838203, "p50_latency_us": 1660.615, "p90_latency_us": 2731.45, "p99_latency_us": 4056.268 }, { "success_rate": 1.0, "operations_per_second": 36011.59288401676, "average_latency_us": 1774.3396349678915, "p50_latency_us": 1668.179, "p90_latency_us": 2704.5480000000002, "p99_latency_us": 3995.2850000000003 }, { "success_rate": 1.0, "operations_per_second": 35952.53787021321, "average_latency_us": 1777.2674508564219, "p50_latency_us": 1667.657, "p90_latency_us": 2735.286, "p99_latency_us": 4046.089 } ], "median": { "success_rate": 1.0, "operations_per_second": 36011.59288401676, "average_latency_us": 1774.3396349678915, "p50_latency_us": 1667.657, "p90_latency_us": 2731.45, "p99_latency_us": 4046.089 } }, "nginx": { "trials": [ { "success_rate": 1.0, "operations_per_second": 49969.83530769965, "average_latency_us": 1278.2637684421481, "p50_latency_us": 1203.4019999999998, "p90_latency_us": 2055.86, "p99_latency_us": 3351.594 }, { "success_rate": 1.0, "operations_per_second": 49436.16079313836, "average_latency_us": 1292.1146184768754, "p50_latency_us": 1206.4479999999999, "p90_latency_us": 2089.914, "p99_latency_us": 3544.8709999999996 }, { "success_rate": 1.0, "operations_per_second": 49378.778193000064, "average_latency_us": 1293.5019977794504, "p50_latency_us": 1214.142, "p90_latency_us": 2078.983, "p99_latency_us": 3267.2360000000003 } ], "median": { "success_rate": 1.0, "operations_per_second": 49436.16079313836, "average_latency_us": 1292.1146184768754, "p50_latency_us": 1206.4479999999999, "p90_latency_us": 2078.983, "p99_latency_us": 3351.594 } } }, "comparison": { "a3s_to_nginx_throughput_ratio": 0.7284463903801991, "a3s_to_nginx_p50_latency_ratio": 1.3822866795750832, "a3s_to_nginx_p90_latency_ratio": 1.3138395071051565, "a3s_to_nginx_p99_latency_ratio": 1.207213343859668, "positions": { "throughput": "nginx_higher", "p50_latency": "nginx_lower", "p90_latency": "nginx_lower", "p99_latency": "nginx_lower" } } }, "openai-stream": { "label": "OpenAI stream", "traffic": "ai-streaming", "unit": "streams/s", "load_generator": "oha", "capability_alignment": "a3s_feature_enabled_vs_nginx_transport", "workload": "Completions JSON stream detection and finite SSE forwarding", "proxies": { "a3s-gateway": { "trials": [ { "success_rate": 1.0, "operations_per_second": 36151.23570517772, "average_latency_us": 1767.574687550793, "p50_latency_us": 1664.61, "p90_latency_us": 2708.344, "p99_latency_us": 4001.5939999999996 }, { "success_rate": 1.0, "operations_per_second": 34670.00941519302, "average_latency_us": 1842.9689122097507, "p50_latency_us": 1727.95, "p90_latency_us": 2835.466, "p99_latency_us": 4172.216 }, { "success_rate": 1.0, "operations_per_second": 35893.128152251964, "average_latency_us": 1780.2640332232465, "p50_latency_us": 1666.7559999999999, "p90_latency_us": 2745.706, "p99_latency_us": 4093.627 } ], "median": { "success_rate": 1.0, "operations_per_second": 35893.128152251964, "average_latency_us": 1780.2640332232465, "p50_latency_us": 1666.7559999999999, "p90_latency_us": 2745.706, "p99_latency_us": 4093.627 } }, "nginx": { "trials": [ { "success_rate": 1.0, "operations_per_second": 49879.31649639336, "average_latency_us": 1280.5620892353486, "p50_latency_us": 1207.921, "p90_latency_us": 2050.849, "p99_latency_us": 3335.625 }, { "success_rate": 1.0, "operations_per_second": 47922.1598893402, "average_latency_us": 1332.9133424958877, "p50_latency_us": 1255.158, "p90_latency_us": 2163.721, "p99_latency_us": 3442.874 }, { "success_rate": 1.0, "operations_per_second": 49558.51932017564, "average_latency_us": 1288.8777465723108, "p50_latency_us": 1202.6409999999998, "p90_latency_us": 2102.147, "p99_latency_us": 3455.919 } ], "median": { "success_rate": 1.0, "operations_per_second": 49558.51932017564, "average_latency_us": 1288.8777465723108, "p50_latency_us": 1207.921, "p90_latency_us": 2102.147, "p99_latency_us": 3442.874 } } }, "comparison": { "a3s_to_nginx_throughput_ratio": 0.7242574767087443, "a3s_to_nginx_p50_latency_ratio": 1.3798551395331315, "a3s_to_nginx_p90_latency_ratio": 1.3061436712085313, "a3s_to_nginx_p99_latency_ratio": 1.189014468725838, "positions": { "throughput": "nginx_higher", "p50_latency": "nginx_lower", "p90_latency": "nginx_lower", "p99_latency": "nginx_lower" } } } }, "proxies": { "a3s-gateway": { "trials": [ { "success_rate": 1.0, "requests_per_second": 38750.361646250894, "average_latency_us": 1649.02694841716, "p50_latency_us": 1549.978, "p90_latency_us": 2521.2479999999996, "p99_latency_us": 3740.229 }, { "success_rate": 1.0, "requests_per_second": 39053.39352193416, "average_latency_us": 1636.1905084363493, "p50_latency_us": 1536.71, "p90_latency_us": 2499.557, "p99_latency_us": 3723.338 }, { "success_rate": 1.0, "requests_per_second": 38359.30408189978, "average_latency_us": 1665.6597540610344, "p50_latency_us": 1567.019, "p90_latency_us": 2541.763, "p99_latency_us": 3812.248 } ], "median": { "success_rate": 1.0, "requests_per_second": 38750.361646250894, "average_latency_us": 1649.02694841716, "p50_latency_us": 1549.978, "p90_latency_us": 2521.2479999999996, "p99_latency_us": 3740.229 } }, "nginx": { "trials": [ { "success_rate": 1.0, "requests_per_second": 52089.64291265093, "average_latency_us": 1226.217796434385, "p50_latency_us": 1147.4479999999999, "p90_latency_us": 1974.038, "p99_latency_us": 3220.202 }, { "success_rate": 1.0, "requests_per_second": 51482.39074350847, "average_latency_us": 1240.771500929106, "p50_latency_us": 1166.453, "p90_latency_us": 2011.157, "p99_latency_us": 3340.4210000000003 }, { "success_rate": 1.0, "requests_per_second": 51389.530609373316, "average_latency_us": 1243.0560159895456, "p50_latency_us": 1172.5749999999998, "p90_latency_us": 2005.2569999999998, "p99_latency_us": 3195.0429999999997 } ], "median": { "success_rate": 1.0, "requests_per_second": 51482.39074350847, "average_latency_us": 1240.771500929106, "p50_latency_us": 1166.453, "p90_latency_us": 2005.2569999999998, "p99_latency_us": 3220.202 } } }, "comparison": { "a3s_to_nginx_throughput_ratio": 0.7526915725283604, "a3s_to_nginx_p50_latency_ratio": 1.3287959309119186, "a3s_to_nginx_p90_latency_ratio": 1.2573191366493173, "a3s_to_nginx_p99_latency_ratio": 1.161488937650495, "positions": { "throughput": "nginx_higher", "p50_latency": "nginx_lower", "p90_latency": "nginx_lower", "p99_latency": "nginx_lower" } }, "limitations": [ "GitHub-hosted runners are shared infrastructure, not a controlled bare-metal lab.", "Each fixture is a small synthetic same-host workload; absolute rates do not predict an upstream-dominated production deployment.", "OpenAI profiles enable Gateway JSON and stream detection while NGINX remains a transport-only baseline, so those rows measure feature cost rather than equivalent policy capability.", "The matrix compares the checked-in A3S release profile with the Ubuntu-packaged NGINX build; it is not a universal product ranking." ] }