Assessing the overhead of offloading compression tasks

Exploring compression is increasingly promising as trade-off between computations and data movement. There are two main reasons: First, the gap between processing speed and I/O continues to grow, and technology trends indicate a continuation of this. Second, performance is determined by energy efficiency, and the overall power consumption is dominated by the consumption of data movements. For these reasons there is already a plethora of related works on compression from various domains. Most recently, a couple of accelerators have been introduced to offload compression tasks from the main processor, for instance by AHA, Intel and Microsoft. Yet, one lacks the understanding of the overhead of compression when offloading tasks. In particular, such offloading is most beneficial for overlap with other tasks, if the associated overhead on the main processor is negligible. This work evaluates the integration costs compared to a solely software-based solution considering multiple compression algorithms. Among others, High Energy Physics data are used as a prime example of big data sources. The results imply that on average the zlib implementation on the accelerator achieves a comparable compression ratio to zlib level 2 on a CPU, while having up to 17 times the throughput and utilizing over 80 % less CPU resources. These results suggest that, given the right orchestration of compression and data movement tasks, the overhead of offloading compression is limited but present. Considering that compression is only a single task of a larger data processing pipeline, this overhead cannot be neglected.