Below are the results of the Elasticsearch nightly benchmarks based on the master branch as of that point in time. The Apache Software Foundation also provides a similar page for the Lucene nightly benchmarks.
On each chart, you can click + drag (vertically or horizontally) to zoom in and then shift + drag to move around. Click on a point to see the full log file. Orange points show failed runs.
This test indexes 8.6M documents (POIs from Geonames, total 2.8 GB json) using 8 client threads and 5000 docs per _bulk request against a single or two nodes running on a dual Xeon X2699 (36 real cores, 72 with hyperthreading), 256 GB RAM and Intel 750 PCIe/NVMe SSD.
Defaults, 2 nodes is append-only, using all default settings, but runs 2 nodes on 1 box (5 shards, 1 replica).
Defaults is append-only, using all default settings.
Defaults (4G heap) is the same as
Defaults except using a 4 GB heap (ES_HEAP_SIZE), because the ES default (-Xmx1g) sometimes hits OOMEs.
Fast is append-only, using 4 GB heap, and these settings:
refresh_interval: 30s index.number_of_shards: 6 index.number_of_replicas: 0 index.translog.flush_threshold_size: 4g
FastUpdate is the same as fast, except we pass in an ID (worst case random UUID) for each document and 25% of the time the ID already exists in the index.