doc: add performance.md

Documentation/docs.md

@@ -24,7 +24,7 @@ Administrators who need to create reliable and scalable key-value stores for the
  - Monitoring
  - [Maintenance][maintenance]
  - [Disaster recovery][recovery]
- - Performance
+ - [Performance][performance]
 
 ## Learning
 
@@ -46,6 +46,7 @@ To learn more about the concepts and internals behind etcd, read the following p
 [glossary]: learning/glossary.md
 [interacting]: dev-guide/interacting_v3.md
 [local_cluster]: dev-guide/local_cluster.md
+[performance]: op-guide/performance.md
 [recovery]: op-guide/recovery.md
 [maintenance]: op-guide/maintenance.md
 [security]: op-guide/security.md

Documentation/op-guide/performance.md

@@ -0,0 +1,13 @@
+# Performance
+
+## Understanding performance
+
+etcd provides stable, sustained high performance. Two factors define performance: latency and throughput. Latency is the time taken to complete an operation. Throughput is the total operations completed within some time period. Usually average latency increases as the overall throughput increases when etcd accepts concurrent client requests. In common cloud environments, like a standard `n-4` on Google Compute Engine (GCE) or a comparable machine type on AWS, a three member etcd cluster finishes a request in less than one millisecond under light load, and can complete more than 30,000 requests per second under heavy load.
+
+etcd uses the Raft consensus algorithm to replicate requests among members and reach agreement. Consensus performance, especially commit latency, is limited by two physical constraints: network IO latency and disk IO latency. The minimum time to finish an etcd request is the network Round Trip Time (RTT) between members, plus the time `fdatasync` requires to commit the data to permanant storage. The RTT within a datacenter may be as long as several hundred microseconds. A typical RTT within the United States is around 50ms, and can be as slow as 400ms between continents. The typical fdatasync latency for a spinning disk is about 10ms. For SSDs, the latency is often lower than 1ms. To increase throughput, etcd batches multiple requests together and submits them to Raft. This batching policy lets etcd attain high throughput despite heavy load.
+
+There are other sub-systems which impact the overall performance of etcd. Each serialized etcd request must run through etcd’s boltdb-backed MVCC storage engine, which usually takes tens of microseconds to finish. Periodically etcd incrementally snapshots its recently applied requests, merging them back with the previous on-disk snapshot. This process may lead to a latency spike. Although this is usually not a problem on SSDs, it may double the observed latency on HDD. Likewise, inflight compactions can impact etcd’s performance. Fortunately, the impact is often insignificant since the compaction is staggered so it does not compete for resources with regular requests. The RPC system, gRPC, gives etcd a well-defined, extensible API, but it also introduces additional latency, especially for local reads.
+
+## Benchmarks
+
+TODO