Benchmarks

This page summarizes curated benchmark results for Eden's Eve gateway runtime. Raw benchmark artifacts are environment-specific and are intentionally not committed. Treat these numbers as a public performance brief from controlled internal runs, not as a universal hardware guarantee.

Redis is the most mature production gateway protocol in this release. Results for AI/LLM, Postgres, Mongo, and other direct gateway surfaces should be read as current direct-path benchmark data and validated against your own workload before capacity commitments.

Summary

Eve is Eden's high-performance data plane. It proxies Redis traffic and OpenAI-compatible AI workloads, and is being evaluated on Postgres gateway traffic, while preserving layer-7 awareness for policy, analysis, routing, mirroring, migration, and telemetry.

The key result from the current benchmark set: Eve is competitive with Envoy on raw gateway throughput while doing substantially more product work inline. On Postgres, Eve outperformed direct Postgres and PgDog on every 512-client workload in the suite. On a matched Redis workload, Eve used about one-third less CPU per completed request than Envoy and produced lower tail latency. Newer direct-path max sweeps identify clean capacity knees for Redis, Postgres, Mongo, and LLM endpoint families.

For capacity planning, start with the Gateway Capacity Planning page. It separates clean operating points from overload peaks.

Capacity Knees

The clean knee is the highest measured load that completed without benchmark errors, Eden shedding, or status failures. Peak completed throughput is useful for diagnosis, but it is not a safe operating target when the same run shows backpressure, client errors, or long queues.

Postgres Gateway

The current Postgres benchmark compared Eve against direct Postgres and PgDog on Adam, Eden's 32-CPU Ubuntu 24.04 benchmark host. The run used pinned CPU groups for the proxy, Postgres, and load generator, 15-second pgbench phases, scale 10, and workloads covering simple reads, prepared reads, autocommit writes, inserts, and mixed read/write traffic. The suite completed 126 benchmark rows with zero failed rows.

At 512 clients, Eve was ahead of both direct Postgres and PgDog on every workload in the suite.

The read-heavy cases show the largest gap, but the write-heavy and mixed cases also improved. The smaller read/write-mix uplift is important context: correctness narrows the amount of work a gateway can safely reorder or multiplex. Eve's Postgres path should take the fast path when protocol and SQL shape make it safe, and return to the conservative path when query or session state requires exact ordering.

These Postgres numbers should be read as development and evaluation data. Redis is still the production-ready gateway protocol in this release.

Redis Gateway

Matched Throughput

At a controlled 400k req/s mixed Redis load where both gateways completed the same work without shedding, Eve matched Envoy throughput with lower latency and lower CPU use.

CPU attribution from the sampled run:

Direct Redis was saturated at this offered load, so it is best read as a saturation signal rather than an added-latency baseline.

Overload Behavior

In a 400k offered open-loop Redis load, Eve sustained materially more completed throughput than Envoy while avoiding shed.

With telemetry export and capture-all request analysis enabled, Eve still held 384k req/s at zero shed with p99 3.93 ms on this workload.

Large Payload Throughput

On 64 KiB Redis GET workloads, Eve was competitive with Envoy on wire throughput and showed cleaner error behavior at saturation.

The ceiling in these runs was the backend, host, and TCP path rather than only gateway worker count.

AI And LLM Gateway

Eve exposes an OpenAI-compatible AI gateway for chat, multi-agent/tool, and streaming workloads. These runs used a synthetic OpenAI-compatible backend and a generic Envoy HTTP proxy.

Sustained Load

Earlier large-response HTTP comparison runs used the same synthetic upstream behind Eve and Envoy. Those runs are useful for wire-throughput behavior, not for the small-response request-rate knee above.

The current small-response direct-path sweep found an 8-vCPU clean knee of 5k req/s and an overload plateau of about 5.6k completed req/s.

Latency

Observability Validation

Eve telemetry was validated end-to-end under live Redis load. In a 30-minute window with a 100k-request Redis workload running at 175,472 req/s, the runtime produced:

• 185 metrics across core, gateway, LLM, IAM, endpoint, workload, migration,

tool-safety, and analytics groups.

• 14,043 gateway metric rows stored in ClickHouse, including request and command

counts, per-command duration, endpoint-vs-overhead timing, parse decode/materialize breakdown, bridge queue/write timing, byte counters, and mirror latency/divergence.

• 2,105 trace spans captured and queryable alongside structured logs.
• Dashboard series coverage from emission through storage, query, and UI.

This matters because Eve's performance profile includes built-in protocol-aware observability rather than only generic proxy counters.

Capability Context

The raw numbers should be read alongside the feature set being exercised. Redis is one measured protocol in this report, but the product surface is broader: Eden is a data and AI gateway for databases, model providers, and agent workloads.

Reproducing Workloads

Source benchmark tools for the Redis and AI workloads live in the repository's [benchmark/](https://github.com/eden-dev-inc/eden/tree/main/benchmark) directory:

• cacophony runs open-loop Redis workloads and synthetic Redis, Postgres, and

Mongo wire backends.

• profile-eve-vcpu-knee-sweep.sh runs Redis vCPU knee and max sweeps.
• profile-postgres-synthetic-vcpu-sweep.sh runs Postgres synthetic gateway

ceiling sweeps.

• profile-mongo-synthetic-vcpu-sweep.sh runs Mongo synthetic gateway ceiling

sweeps.

• profile-llm-synthetic-vcpu-sweep.sh runs OpenAI-compatible LLM request-rate

sweeps.

• ai-workload runs synthetic OpenAI-compatible HTTP workloads.

The Postgres results above used a pgbench suite against direct Postgres, PgDog, and Eve on Adam with pinned CPU groups. The curated result table is published here; raw host artifacts and scenario files are not committed because they are environment-specific.

Keep raw result files, profiler captures, and host-specific scenario files outside the repository unless they are deliberately curated for publication.