We release our simulator(simulation/) and prototype(controller/ and netfpga/) code here. This OpenSketch platform currently has heavy hitters and superspreaders implementation as examples built upon several sketch-based building blocks we discussed in the paper. Many other measurement tasks can be implemented on top of it with simple controller code and limited data plane resources, because it only need to configure existing building blocks and analyze the collected data to report the results. We will add more examples for future releases.
We are willing to hear your feedback. Please let us know if you'd like to build other measurement tasks on top of opensketch. We are really interested in getting feedbacks from you on the requirements on opensketch library and data plane, and providing support for your applications. It's also much appreciated if you'd like to contribute to the opensketch libarary and applications.
If you have further questions, feel free to contact us. rmiao@usc.edu, lavanyaj@cs.stanford.edu, minlanyu@usc.edu
This is a packet-trace(CAIDA packet trace) driven simulator we built to compare with NetFlow and other streaming algorithms. From our evalution, it provides a better memory-accuracy tradeoff than NetFlow and achieves comparable accuracy to the streaming algorithms, while provindg generality.
This is control plane code for our prototype. Each sketch has two functions for the measurement program to use:
(1)configure : to specify the packet fields, the memory constraint, and the number of hash functions to use
(2)query : to periodically get the statistics
This two basic functions also defines two types of APIs, including set Hash function, constrain sram counter size, and dump sram counter, etc. The controller employs PCI interface to communicate with data plane, which reads or writes the hardware registers through IOCTL calls.
The measurement program in the controller can periodically query the sketches about the statistics. According to the received statistics, the measurement programs may install new sketches or change the accuracy requirements accordingly. The sketches automatically queries the data plane to get the counters in order to generate the right statistics to the measurement program.
This is opensketch data plane prototype in netfpga, including Header Parser, Hashing, Wildcard Lookup, and SRAM Counter in a pipeline. Since we simply pull packet headers to collect the statistics without changing the packet bus, there is no effect on packet forwarding latency and throughput. As a packet enters, the Header Parser pulls the related fields from the packet header, which are then hashed by hash functions in parallel. We then pass the packet header and hash values to the Wildcard Lookup, where we implement wildcard rule matching in parallel. For each matched rule, we update the corresponding counters in the SRAM Counter. We use the entire 4.5 MB on-board SRAM to store the counters.
This OpenSketch modules should be inserted into the reference switch pipeline and then compiled following NetFPGA developer guide (http://wiki.netfpga.org/foswiki/bin/view/NetFPGA/OneGig/DevelopersGuide).