Previous Section - FireHose WWW Site - Next Section


Streaming data arrives continuously and in volumes and rates that are ever increasing. Timely processing of streaming data is computationally challenging due to limited resources. These include limited CPU operations that can be performed on a datum before the next one arrives, limited memory for storing state information about the stream, limited disk storage for the data itself so that the datums may only be seen once, and limited budgets for energy consumption or CPU/memory/disk resources.

In some settings, it's not known in advance what interesting features are embedded in the stream of data. In an algorithmic sense, this makes the above limitations even more challenging. In addition, the hardware options for collecting streaming data are increasing in variety and capability from multicore CPUs, to GPUs, to special-purpose processors like Tilera chips, to solid-state (Flash) disks. While choices are a good thing, they also challenge developers of stream-processing software to support a growing range of hardware.

Computing on streaming data has been referred to as "drinking from a firehose". Hence the name FireHose for this suite of stream processing benchmarks.

The main purpose of FireHose is to enable comparison of streaming software and hardware, both quantitatively vis-a-vis the rate at which they can process data, and qualitatively by judging the effort involved to implement and run the benchmarks. Because we are both users and designers of software for stream processing, we have wanted for some time to define useful metrics for the speed and flexibility various software/hardware solutions can offer. FireHose represents our initial effort in this regard.

We plan to release the benchmark definitions and sample implementation code openly on the FireHose web site, so that anyone can implement them with their own stream-processing software and/or run them on their own hardware. Our hope is that this effort will generate interest and feedback from the academic and commercial streaming communities. With their help in refining and exanding the suite of benchmarks, and their contributions of performance results, we hope the FireHose benchmarks can become a valuable resource for those interested in stream processing.

The initial benchmarks in the FireHose suite reflect common tasks typical of real-time processing of network packets arriving at high rates as a network is monitored, e.g. for cybersecurity purposes. We plan to augment the suite with benchmarks that process streaming graph edges, and are open to additional suggestions as well.