Tuning AutoSys Agents for High-Latency and High-Burst Throughput

Products

• Autosys Workload Automation
• Workload Automation Agent

Issue/Introduction

During load testing of AutoSys Agents in high-latency regions (e.g., cross-continental links), concurrent job bursts may execute in "chunks" (e.g., blocks of 40 jobs) rather than starting simultaneously.
This results in a staggered start where a large queue (e.g., 1,000 jobs) may take several minutes to fully transition to a "Running" status despite 100% execution success.

Environment

• Product: AutoSys Workload Automation
• Component: System Agent
• Operating System: Rocky Linux (or other Linux distributions)
• Network: High-latency WAN (e.g., North America to China)

Cause

The chunking behavior is typically caused by default thread pool limits in the agent or operating system socket backlogs being reached while waiting for network acknowledgments over high-latency links (e.g., ~4 seconds per transaction).
This creates backpressure and exhausts available resources during extreme "burst" scenarios.

Resolution

To improve burst throughput and reduce chunking, tune the following parameters at both the Agent and Operating System levels.

1. AutoSys Agent Configurations (agentparm.txt)

• Add or modify the following parameter

communication.receiver.pool.size=10
communication.transmitter.pool.size=10

• Restart the agent to apply the changes.

2. Operating System Configurations (Rocky Linux)

Note: These changes must be reviewed and implemented by your System Administrators.

• User Resource Limits (ulimit): Ensure the agent user has sufficient limits for processes and file descriptors.
  • ulimit -u (max user processes): Set to >2048 (or higher for extreme volumes).
  • ulimit -n (open files): Set to >2048 to support the necessary network sockets and log files.
• TCP Listen Backlog (net.core.somaxconn): If the manager sends a massive burst of concurrent connection requests, a low OS backlog (default is often 128) may drop packets. Increase this value in /etc/sysctl.conf to handle larger SYN bursts.
• TCP Window Scaling: Verify that net.ipv4.tcp_window_scaling = 1 is enabled. This is mandatory for maintaining high throughput on high-latency links by allowing TCP windows to scale past 64KB.

 

 

 

Additional Information

For Transmitter side of things, see This Article