> ## Documentation Index
> Fetch the complete documentation index at: https://docs.lasso.sh/llms.txt
> Use this file to discover all available pages before exploring further.

# Routing overhead benchmark

> Measured latency Lasso's routing engine adds at 10,000 requests per second, and what the result does and doesn't show

# Routing overhead benchmark

Lasso measured how much latency its routing engine adds in front of an
upstream: under 1.1 ms at p95, at 10,000 requests per second, across five
isolated trials. Use this result to bound Lasso's own cost on a request path;
your provider's latency and your network dominate the rest.

| Result | Value |
| - | - |
| Median added p95 across trials | 0.95 ms |
| Added p95, all five trials | 0.70 to 1.08 ms |
| Routed requests successful | 750,000 of 750,000 |

The full report, raw results and harness revision are on
[lasso.sh/benchmarks/routing-overhead.html](https://lasso.sh/benchmarks/routing-overhead.html).

## How it was measured

| Trial | Direct p95 | Routed p95 | Added p95 |
| - | - | - | - |
| 1 | 5.69 ms | 6.64 ms | 0.95 ms |
| 2 | 5.74 ms | 6.44 ms | 0.70 ms |
| 3 | 5.61 ms | 6.69 ms | 1.08 ms |
| 4 | 5.73 ms | 6.53 ms | 0.81 ms |
| 5 | 5.73 ms | 6.70 ms | 0.97 ms |

* Each trial sent 150,000 `eth_getBalance` requests over HTTP in 15 seconds,
  at fixed arrivals of 10,000 per second, after a 5-second warmup, through the
  `fastest` strategy to a synthetic upstream that answers in 1 ms.
* "Added p95" is routed p95 minus the paired direct-to-upstream p95, measured
  from each request's intended arrival time.
* Lasso ran on four BEAM schedulers pinned to four CPUs in a fresh container
  per trial, on a Linux/arm64 Docker host. Load generators and upstreams had
  their own CPUs.
* A separate 60-second instrumented run completed 600,000 of 600,000 requests
  with 2.29 ms added p95.

## Limits

* This is routing-engine overhead, not a hosted-capacity promise. It excludes
  internet ingress, authentication, metering, multiple regions, WebSocket
  subscriptions and real providers.
* One method against one controlled upstream. Other methods, payload sizes,
  retries and deployment shapes can differ.
* A difference of two p95 values is a paired distribution comparison, not a
  per-request overhead.

## Next

* [Route latency-sensitive reads](/cloud/latency-sensitive-rpc-reads): measure
  your own path.
* [Routing](/cloud/routing): what the engine does with that time.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.