Applications running on Network on Chip (NoC) based multicore systems demand increased on-chip network bandwidth that can cater to the need for intensive communication among the cores and caches. Due to strict area and power budget, the bandwidth offered by NoC is very limited. Data-intensive and communication-centric applications on encountering a cache miss lead to a considerable burden on the underlying network for transferring blocks from multiple cache hierarchies to the requesting core as packets. This increases the packet transmission latency, thereby slowing down the system performance. Also, NoC being the highest component of power consumption after the cores, an increase in packets increases the dynamic power consumption of NoC. The article proposes FlitZip that addresses the problem by reducing on-chip traffic through compressing network packets. Hence, the compressed packet requires less bandwidth during its transfer, reducing the network's power consumption. Experimental analysis shows that FlitZip achieves a better compression ratio of 52 percent, reduces packet latency and bandwidth utilization by 19.28 and 27 percent, respectively. It also reduces the area and power consumption of the de/compression units by 53.33 and 62.3 percent, respectively, compared to the state-of-the-art packet compression technique, NoΔ.

中文翻译：

---

FlitZip：多处理器片上系统中 NoC 的有效数据包压缩


在基于片上网络 (NoC) 的多核系统上运行的应用程序需要增加片上网络带宽，以满足内核和缓存之间密集通信的需求。由于严格的面积和功率预算，NoC提供的带宽非常有限。数据密集型和以通信为中心的应用程序在遇到缓存未命中时，会导致底层网络承受相当大的负担，因为底层网络需要将块作为数据包从多个缓存层次结构传输到请求核心。这会增加数据包传输延迟，从而降低系统性能。此外，NoC 是继内核之后功耗最高的组件，数据包的增加会增加 NoC 的动态功耗。该文章提出了 FlitZip，通过压缩网络数据包减少片上流量来解决该问题。因此，压缩数据包在传输过程中需要更少的带宽，从而降低了网络的功耗。实验分析表明，FlitZip 实现了 52% 的更好压缩率，数据包延迟和带宽利用率分别降低了 19.28% 和 27%。与最先进的数据包压缩技术 NoΔ 相比，它还将解/压缩单元的面积和功耗分别减少了 53.33% 和 62.3%。