Add benchmark numbers to README (#5)

… and slightly tweak the benchmark to provide more relevant
measurements for throughput.

README.md

@@ -110,11 +110,15 @@ Did we satisfy those requirements?
   * [`harmonia`](https://github.com/helsinki-systems/harmonia) - A Rust rewrite
     of `nix-serve`
 
+  See the Benchmarks section below for more details
+
 * Backwards-compatibility
 
-  We have excellent backwards-compatibility.  In particular, in the vast
-  majority of cases, you can simply replace `pkgs.nix-serve` with
-  `pkgs.nix-serve-ng` and make no other changes.
+  We have excellent backwards-compatibility with one major exception:
+  `nix-serve-ng` does not support MacOS (whereas `nix-serve` does).
+
+  Other than that, in the vast majority of cases, you can simply replace
+  `pkgs.nix-serve` with `pkgs.nix-serve-ng` and make no other changes.
 
   * Our executable shares the same name (`nix-serve`) as the original program
 
@@ -173,3 +177,52 @@ Did we satisfy those requirements?
   You don't need to define or use any new NixOS options.  You continue to use
   the old `services.nix-serve` options hierarchy to configure the upgraded
   service.
+
+## Benchmarks
+
+The test environment is a large server machine:
+
+* CPU: 24 × Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz
+* RAM: 384 GB (24 × 16 GB @ 2133 MT/s)
+* Disk (`/nix/store`): ≈4 TB SSD
+
+Legend:
+
+* **Fetch present NAR info ×10**: Time to fetch the NAR info for 10 different files that are present
+* **Fetch absent NAR info ×1**: Time to fetch the NAR info a single file that is absent
+* **Fetch empty NAR ×10**: Time to fetch the NAR for the same empty file 10 times
+* **Fetch 10 MB NAR ×10**: Time to fetch the NAR for the same 10 MB file 10 times
+
+Raw numbers:
+
+| Benchmark                  | `nix-serve`      | `eris`           | `harmonia`       | `nix-serve-ng`   |
+|----------------------------|------------------|------------------|------------------|------------------|
+| Fetch present NAR info ×10 | 2.09 ms ± 66  μs | 41.5 ms ± 426 μs | 1.57 ms ±  91 μs | 1.32 ms ±  33 μs |
+| Fetch absent NAR info ×1   | 212  μs ± 18  μs | 3.42 ms ± 113 μs | 139  μs ±  11 μs | 115  μs ± 6.2 μs |
+| Fetch empty NAR ×10        | 164  ms ± 8.5 ms | 246  ms ±  20 ms | 279  ms ±  10 ms | 5.16 ms ± 368 μs |
+| Fetch 10 MB NAR ×10        | 291  ms ± 8.7 ms | 453  ms ±  19 ms | 487  ms ±  41 ms | 86.9 ms ± 3.0 ms |
+
+Speedups (compared to `nix-serve`):
+
+| Benchmark                  | `nix-serve`      | `eris`           | `harmonia`       | `nix-serve-ng`   |
+|----------------------------|------------------|------------------|------------------|------------------|
+| Fetch present NAR info ×10 | 1.0              | 0.05             | 1.33             |  1.58            |
+| Fetch absent NAR info ×1   | 1.0              | 0.06             | 1.53             |  1.84            |
+| Fetch empty NAR ×10        | 1.0              | 0.67             | 0.59             | 31.80            |
+| Fetch 10 MB NAR ×10        | 1.0              | 0.64             | 0.60             |  3.35            |
+
+We can summarize `nix-serve-ng`'s performance like this:
+
+* Time to handle a NAR info request: ≈ 100 μs 
+* Time to serve a NAR: ≈ 500 μs + 800 μs / MB
+
+You can reproduce these benchmarks using the benchmark suite.  See the
+instructions in [`./benchmark/Main.hs`](./benchmark/Main.hs) for running your
+own benchmarks.
+
+Caveats:
+
+* We haven't used any of these services' tuning options, including:
+  * Tuning garbage collection (for `nix-serve-ng`)
+  * Tuning concurrency/parallelism/workers
+* We haven't benchmarked memory utilization