= yaml = title: Vanadium Performance toc: true = yaml =
Performance of the Vanadium APIs is measured with microbenchmarks. benchmarks.v.io records these results at various snapshots of the codebase and on multiple platforms. Since the total number of benchmarks served by benchmarks.v.io is somewhat overwhelming, this page lists out a small subset that can be considered representative of Vanadium.
The numbers below are for using the Go API. Benchmarks for other languages (Java in particular) are not integrated into the flow of continuous measurement yet, but are intended to be.
All the numbers are currently one click away. We hope to restructure this page to embed live results to avoid that one click in the future.
In all the benchmarks here, network round-trip time is zero. Thus, when RPCs are executed ‘in the wild’, network round-trip time must be added.
Results: v.io/x/ref/runtime/internal/rpc/benchmark.BenchmarkConnectionEstablishment
Establishment of a mutually authenticated, confidential network connection between two processes. This includes the time it takes to establish a TCP connection and execute the Vanadium authentication protocol over it.
Results: v.io/x/ref/runtime/internal/rpc/benchmark.Benchmark____1B
Sending a 1 byte “echo” request and receiving the response over an established connection. The Vanadium RPC protocol multiplexes RPCs over a single established connection.
Results: v.io/v23/sycnbase/nosql.BenchmarkTinyPut.
Writing a small piece of structured information to the syncbase daemon via the Table API.
Results: v.io/v23/sycnbase/nosql.BenchmarkTinyGet.
Reading read a small piece of structured information from the syncbase daemon via the Table API.
Results: v.io/v23/syncbase/featuretests.BenchmarkPingPongPair.
As of February 2016, this measures 500x the time it takes for peers in a syncgroup to notice updates made to each other. For example, on desktop/laptop class machines benchmarks.v.io reported 1m40s, which means 100/500 seconds = 200ms to sync an update. The 500x multiplier is an artifact of the benchmark implementation, not the sync protocol.
This also includes any idle time between attempts to contact the peer syncbase instance for updates. Changes to the sync protocol (changing the polling interval or using a push-notification mechanism) were being iterated on at the time of this writing. These numbers are very sensitive to such changes.