2.11. Benchmarking and Performance planning

2.11.1. Performance planning - the Theory

ARender’s architectural design is split into two distinct components

  • Presentation Server (HMI)
  • Rendition Server

2.11.1.1. Presentation Server

The Presentation Server, deployed in a standard J2EE Application Server, connects the user, the ECM repository and the Rendition Server. As such it does not perform real work, and does not involve much CPU or RAM in order to work. Standard recommandations for these servers involve at least 128Mb of extra RAM of the JVM in the Application Server.

2.11.1.2. Rendition Server

The Rendition Server performs three distinct, important tasks :

  1. Identify document type, analyze its layout (number of pages, page size)
  2. Extract text information for each page, for search/annotation/copy-paste features
  3. Render pages into images in a streaming manner

Document identification and analysis involves mostly disk I/O and a bit of RAM, but the real worker here is the page rendition part, which requires a large part of RAM and CPU power. As a result, correctly dimensioning a production-grade ARender platform only implies to have a consistent estimation of the number of pages to be rendered per second.

Per-page performance

ARender has a very powerful page image renderer, written in native C for performance. As a rule of thumb, modern hardware and virtual machines are able to render pages with the following performances :

Image size Width (pixels) Timing (ms)
Thumbnails 100px 23ms
Mobile device screen 348px 38ms
Desktop screen, 2/3 size 595px 68ms
Desktop screen, full screen 1280px 191ms

These figures are taken at server-side (not considering network time), on a 8-CPU multi-document scenario. Collecting these figures allows to compute how much CPUs are required to process the number of pages asked.

Estimated usage

On the other hand, we have to estimate the number of pages the rendition server will ask. This is usually estimated by estimating the following :

  • Total number of users (ECM users for example)
  • Average percentage of users simultaneously active
  • Average number of documents opened per hour, per user
  • Average number of pages per document, per document format

Having this allows to give an accurate average for the number of documents open per second, and then the number of pages open per second. Final dimensioning should take care of the fact that :

  • ARender does not necessarily render each page of the document : only the pages shown to the user (and the ones surrounding) are rendered. As a result, the larger the document, the smaller the ratio between pages rendered and total pages.
  • Dimensioning must be done for the busy-hour (peak activity), so whole-day averages can be misleading. As a rule of thumb, a busy-hour can represent from 25% to 50% of the daily activity, depending on the business and organisation.

2.11.2. Clustering and Rendition farms

ARender natively supports horizontal scalability, both for the Presentation server (using standard J2EE techniques), and for the Rendition server (using ARender’s rendition-client features).

Horizontal scaling increases both reliability (high-availability) and performance (load balancing). Because load balancing is performed on a per-document basis without the need of node synchronization, two Rendition servers are considered to open twice as much documents as a single Rendition Server, with no performance loss.

2.11.3. Benchmarks - Performance in the field

Company type - Usage Total number of users Average number of documents opened daily Number & type of Rendition Servers
Insurance company - Agencies & Extranet 20,000+ 350,000 2x 16 CPU, Physical machine
Real Estate company - Internal/Extranet usage 1,500 20,000 2x 4 CPU, Virtual machine
Insurance company - Internal usage + Batch-based document conversion using ARender 800 60,000 6x 4 CPU, Virtual machine