Processor Impact on Agent Performance

Processors are a very powerful tool in BindPlane OP. They allow you to transform your data before it arrives at your destination. This can be useful for adding attributes, filtering, and converting logs to metrics. The question is, how do processors impact agent performance?

We'll look at some common uses cases for processors and see how they can impact agent performance.

Table of Contents

Benchmarking Setup

For our benchmarking setup we have two agents running on a GCP n2-standard-2 Compute Engine instance, which has 2 vCPUs and 8GB of memory. One agent will be the test agent that we will be applying processors to. The other agent will be acting as our destination. It is configured with an OTLP Source and custom destination to log the output of the data it receives.

For the log generation tests static JSON logs were generated at a rate of 1000 logs per second, across 10 files. This roughly equates to 140MiB of logs per minute. The log source is the Filelog source with JSON parsing configured.

For metric generation the Host Metrics source was used with the default configuration.

Control Runs

A one hour control run was done for both the log and metric generation tests. This was to establish a baseline for the agent's performance without any processors applied.

Logs Control

The control run for logs was conducted collecting from 10 log files with logs generating at 1000 logs per second to each file. This results in a throughput of ~140MiB/m. The agent's average CPU usage was 27% and average memory usage was 2.8% (~230MB).

This is the topology view of the control configuration:

Logs Control Topology

These are the graphs showing CPU and memory usage over the hour long run:

Logs Control Results

Metrics Control

The control run for metrics had a data throughput of ~99KiB/m. The agent's average CPU usage was 0.2% and average memory usage was 2.1% (~16MB).

This is the topology view of the control configuration:

Logs Control Topology

These are the graphs showing CPU and memory usage over the hour long run:

Logs Control Results

Deduplicating Logs

Deduplicating logs is a common data reduction use case for logs. It can help reduce the number of redundant logs that are sent to a destination.

For our tests we used the Deduplicate Logs processor to deduplicate logs. The processor is configured to duplicate logs excluding the timestamp and log.file.name attributes.

The benchmark was run for 30 minutes. The agent's average CPU usage was 16% and average memory usage was 2.8% (~230MB). This is actually lower than the control run. This is due to the fact that the logs are being deduplicated and the agent is processing less data.

This is the topology view of the benchmark configuration:

Logs DeDup Topology

These are the graphs showing CPU and memory usage over the 30 minute run: (The benchmark takes place in the section between the blue lines.)

Logs DeDup Results

Benchmark Conclusion

The above images show that apply a log deduplication processor does not negatively impact the agent's performance. It actually reduces CPU usage as it reduces the amount of data the agent has to process.

Converting Logs to Metrics

Converting logs to metrics is a great way to reduce cost and data ingestion of logs if you just want to see some metric based off of the logs.

For this benchmark we will be solving the use case of wanting to know the number of logs for each severity and discard the logs themselves. We will use the Count Telemetry processor to count the number of info, warning, and error logs. We will use the Filter by Field processor to filter out the logs and only send the metrics.

Here is the order in which the processors are applied:

Logs To Metric Processors

The benchmark was run for 30 minutes. The agent's average CPU usage was 20% and average memory usage was 2.5% (~205MB). Like the Log DeDuplication Benchmark, this is actually lower than the control run.

This is the topology view of the benchmark configuration: (It shows 0B/m after the processors as they are all being filtered out.)

Logs To Metrics Topology

These are the graphs showing CPU and memory usage over the 30 minute run: (The benchmark takes place in the section between the blue lines.)

Logs to Metrics Results

Benchmark Conclusion

The above images show that apply even with four processors in use it does not negatively impact the agent's performance. Both CPU and memory usage are lower than the control run.

Lookup Fields

The Lookup Fields processor allows adding additional fields to telemetry based on existing fields. This can be useful for adding additional context before sending to a destination.

This benchmark we add an additional field to the logs based on the env field in the body.

The benchmark was run for 30 minutes. The agent's average CPU usage was 25% and average memory usage was 2.5% (~205MB). This is fairly close to the control run.

This is the topology view of the benchmark configuration:

Logs Lookup Topology

These are the graphs showing CPU and memory usage over the 30 minute run: (The benchmark takes place in the section between the blue lines.)

Logs Lookup Results

Benchmark Conclusion

The results are fairly close to the control run. We don't see a performance improvement like the previous runs as we are adding a small amount of metadata to the logs rather than reducing overall throughput.

Delete Fields

The Delete Fields processor allows removing fields from telemetry. This can be useful for removing extraneous or sensitive data from telemetry before sending it to a destination.

This benchmark was run against both logs and metrics deleting fields from both.

Logs

The benchmark was run for 30 minutes. The agent's average CPU usage was 24% and average memory usage was 2.4% (~196MB). This is fairly close to the control run.

This is the topology view of the benchmark configuration:

Delete Fields Log Topology

These are the graphs showing CPU and memory usage over the 30 minute run: (The benchmark takes place in the section between the blue lines.)

Delete Fields Log Results

Benchmark Conclusion

The results are fairly close to the control run with using slightly less CPU and memory as the logs are slightly smaller due to the fields being removed.

Metrics

The benchmark was run for 30 minutes. The agent's average CPU usage was 0.2% and average memory usage was 2.1% (~172MB). This is identical to the metrics control run.

This is the topology view of the benchmark configuration:

Delete Fields Metrics Topology

These are the graphs showing CPU and memory usage over the 30 minute run: (The benchmark takes place in the section between the blue lines.)

Delete Fields Metrics Results

Benchmark Conclusion

The results were identical to the metrics control run. Since volume of metrics and the rate at which they are processed is low, the impact of the processor is negligible.

Takeaways

The benchmarks show that using processors do not negatively impact the agent's performance. In fact, in some cases it can reduce the amount of data the agent has to process and reduce CPU usage.

Data reduction can significantly reduce the resource usage of an agent. Ordering processors so data reduction happens earlier in the pipeline can help reduce the overall resource usage of the agent.

The impact of a processor is dependent on the volume of data being processed. The log benchmarks showed a greater impact from processors than the metric benchmarks. This is due to the volume of logs being processed being much higher than the volume of metrics.