Abstract
Network processing plays an important role in the development of Internet as more and more complicated applications are deployed throughout the network. With the advent of new platforms such as network processors (NPs) that incorporate novel architectures to speedup packet processing, there is an increasing need for an efficient method to facilitate the study of their performance. In this paper, we present a tool called SimNP, which provides a flexible platform for the simulation of a network processing system in order to provide information for workload characterization, architecture development, and application implementation. The simulator models several architectural features that are commonly employed by NPs, including multiple processing engines (PEs), integrated network interface and memory controller, and hardware accelerators. ARM instruction set is emulated and a simple memory model is provided so that applications implemented in high level programming language such as C can be easily compiled into an executable binary using a common compiler like gcc. Moreover, new features or new modules can also be easily added into this simulator. Experiments have shown that our simulator provides abundant information for the study of network processing systems.
Introduction
The design and evaluation of Network Processors (NPs) require a delicate balance between high-speed custom hardware performance and the flexibility of programmable CPUs. Existing simulation tools often fall short: General-Purpose Processor (GPP) simulators like SimpleScalar focus on desktop CPU features (e.g., out-of-order execution) that are irrelevant for NPs, leading to slow and inaccurate simulations. Conversely, vendor-specific simulators are often "black boxes" that are difficult to modify and require low-level microcode programming.
SimNP addresses these challenges by providing a modular, execution-driven simulator designed specifically for NPs. It emulates the ARM instruction set and models specialized hardware like TCAMs and hardware accelerators, allowing researchers to evaluate complex networking workloads using standard C code.
System Architecture and Methods
SimNP is built upon the software architecture of the SimpleScalar/ARM toolset. Its modular design allows for easy expansion and modification of architectural components.
Key technical features include:
- Processing Engines (PEs): Supports up to 32 PEs emulating the ARM7 integer instruction set, enabling the use of standard
gccfor compilation. - Execution-Driven Simulation: Uses a front-end for instruction interpretation and a back-end performance model for timing, bus contention, and queuing delays.
- Specialized Memory Modeling: Includes models for DRAM, SRAM, and TCAM (Ternary Content Addressable Memory), the latter being critical for high-speed Longest Prefix Matching.
- Hardware Accelerators: Features "clusters" for offloading tasks. Cluster A handles data-intensive tasks via DMA, while Cluster B focuses on header-based tasks.
- Memory-Mapped I/O: Hardware modules are accessed via specific memory addresses, facilitating the addition of new components without altering the core ISA.
Experimental Setup
The platform was evaluated using real-world packet traces from NLANR and a 127,000-entry AT&T East route table for IP forwarding tests. Synthetic traffic was also utilized for stress-testing specific hardware modules.
Key Results and Insights
The simulation results highlight critical performance characteristics of NP architectures. A primary finding is that the bottleneck in NPs often shifts from raw CPU speed to the System Bus and Memory Bandwidth.
Multi-core Efficiency and Scaling
Different networking tasks exhibit varied scaling behaviors:
- Packet Fragmentation: This DRAM-heavy task hits a performance wall quickly. As shown in the figure above, stall cycles increase exponentially from ~70 million with 1 PE to over 1,000 million with 32 PEs, indicating that adding cores without improving I/O is counterproductive.
- LC-Trie (IP Lookup): This SRAM-heavy task scales more efficiently. The "sweet spot" for the simulated architecture was identified at 8 PEs. Beyond this point, performance gains are marginal due to increased overhead and suspension states.
Impact of Hardware Accelerators
Implementing the AES (Advanced Encryption Standard) algorithm as a hardware accelerator provided deterministic processing times and significantly lower cycle counts compared to software-only implementations. This underscores the importance of specialized hardware for data-intensive networking tasks.
Conclusion and Future Work
SimNP provides a robust and flexible platform for NP research, emphasizing ease of programming through ARM ISA support. Future work will focus on enhancing the processing engine's cycle accuracy, implementing a complex cache hierarchy with coherence protocols, and integrating real-time debugging tools to further streamline application verification.
Summary updated