News / 28 November 2016

Performance in the iRODS pluggable rule engine framework


iRODS 4.2.0 introduces the pluggable rule engine framework. The pluggable rule engine framework allows iRODS rules to be written in languages other than the iRODS rule language. For example, iRODS 4.2.0 ships with a compiled default policy plugin (written in C++), and a Python rule engine plugin will be released shortly thereafter. Adding support for additional languages is also straightforward, as one only has to write the plugin (<1000 LOC), add the shared object to her iRODS installation, and update server_config.json accordingly.

The cost of flexibility

The pluggable rule engine framework represents a huge upgrade in flexibility from iRODS versions 4.1.9 and previous, in which rules were required to be written in the iRODS rule language. However, this impressive power came at a cost, as rule invocation was between 1.6 and 1.7 times more expensive in the June 2016 branch of iRODS 4.2 compared to iRODS 4.1.9.

Fast and flexible

Despite the additional power and flexibility offered by the pluggable rule engine framework, a two-thirds performance penalty on rule invocation is unacceptable. Our goal then became keeping the flexibility of the pluggable rule engine framework while simultaneously improving performance to match or improve upon iRODS 4.1.9.

June 2016

(NOTE: All speed comparisons are based on timing of running the ils icommand 1000x on a fresh install of iRODS. This was repeated 5 times, and the median time was chosen as the timing value for each configuration.)

Table 1 shows the results of early performance testing between 4.1.9 and 4.2. FORK indicates that rather than fork-and-exec’ing the irodsAgent as in versions 4.1.9 and previous, the agents were just forked irodsServer processes that called the runIrodsAgent function. !DYNAMIC_PEP indicates that the dynamic pre- and post-operation policy enforcement points that are being introduced in iRODS 4.2 were disabled for testing. !ENABLE_RE indicates that the pluggable rule engine was turned off by removing the ENABLE_RE compile flag.

version relative timing
4.1.9 1.0x
4.2 (FORK & !ENABLE_RE) 1.04x
4.2 (FORK & !DYNAMIC_PEP) 1.10x
4.2 (FORK) 1.16x
4.2 1.67x

Table 1 - Performance comparison between iRODS 4.1.9 and iRODS 4.2 (June 2016)

Changing to fork rather than fork-and-exec to spawn agents consistently represented the single biggest performance improvement. Loading shared libraries was observed to cost approximately 1/3 of the total cost of rule invocation. Disabling part (the dynamic PEPs) or all (via ENABLE_RE) of the rule engine plugin framework further improved performance to the point that it was comparable to 4.1.9, but at the cost of new features intended for 4.2.

October 2016

There were three main design decisions that enabled us to introduce the pluggable rule engine framework in iRODS 4.2 as intended, while achieving our goal of rule invocation performance comparable to that of iRODS 4.1.9.

Starting agents with fork instead of fork-exec

Performance analysis of the iRODS 4.2 (June 2016) code revealed that a substantial amount of time (\~30%) was being spent in loading shared libraries at the start of each agent process. We sought to avoid this process by forking agents off of the running iRODS server process rather than fork-and-exec’ing a separate executable. Doing so required a refactoring of the code to add a second long-running process responsible for creating the agent processes that handle each rule call, but enabled us to pay the cost of shared library loading only once, at server start, rather than at the creation of each agent process.

Moving the dispatcher into the rule engine framework

In the iRODS 4.2 (June 2016) code, there was a separate rule engine plugin to handle dispatching of rule calls to the other rule engine plugins. This plugin was required even in the case where there was only one rule engine plugin installed. This imposed a performance penalty, as rule arguments needed to be packed and unpacked an additional time for every rule call. To avoid this overhead, we moved the functionality of the dispatcher plugin back into the core iRODS codebase.

Implementing default policy in C++

For many users and applications, the default policy options in iRODS are sufficient. In this case, there is no need to incur the performance overhead of interpreting rules written in the iRODS rule language at runtime. Instead, we can write these rules in C++ and compile them, improving performance and reducing potential confusion associated with the core.re rulebase.

Conclusions

By forking the irodsAgent processes rather than fork-and-exec'ing them, moving the dispatcher rule engine plugin back into the iRODS codebase, and implementing the policy defaults in C++ rather than in the iRODS rule language, we were able to improve performance in the 4.2 engineering preview to be comparable to performance in iRODS 4.1.9 and still introduce the pluggable rule engine framework. Table 2 shows the results of performance testing of the 4.2 engineering preview versus version 4.1.9.

version timing relative timing
4.1.9 104.57s 1.0x
4.2 (FORK, DISPATCHER, & DEFAULT) 106.88s 1.02x
4.2 (FORK & DISPATCHER) 123.50s 1.18x
4.2 (FORK) 132.47s 1.27x
4.2 168.49s 1.61x

Table 2 - Performance comparison between iRODS 4.1.9 and iRODS 4.2 engineering preview


Rick Skarbez