The first part of this post is by way of thanks to the Newton Fund , the UK academies (say, The Royal Society) and the Mexican Academy of Sciences (AMC) for a Newton mobility grant which will allow me to visit my colleague Armando Castaneda at UNAM for six weeks in August and September this year. The call funds upto three months of ‘foreign activity’ so I could have possibly asked for more time but I was unsure of being able to get away for so long. Armando managed to visit me at Belfast some time ago so this can be even thought of as a return visit! He even managed to time his visit to coincide with the Belfast Marathon (while he’s training for a marathon) – `curiosier and curioser’, as Alice would say!
So, what’s this resilient compact routing? About three years ago, I was fortunate to be an I-CORE postdoc with Danny Dolev and Armando was a postdoc at Technion. We started discussing ideas around routing and possibly due to my long line of work with self-healing algorithms (on which many blog posts may follow!) we started to gravitate towards the question: Can we route messages despite failures in the network? At about the same time, Shiri Chechik bested our Leader Election paper (On the complexity of universal leader election) with her paper Compact routing schemes with improved stretch for the best paper award at PODC 2013. With many life-changing events in-between (such as getting faculty positions and moving to many degrees drop in average temperature and many degrees more of precipitation!), the first paper in this line has just managed to struggle over in early 2016. Compact Routing messages in self-healing trees (Arxiv) was a finalist for the best paper award in ICDCN 2016.
So, what’s this resilient compact routing (take 2)? Routing is a very important `primitive’ for networks – the ability of the network to take a message from a source node and deliver it to a target. We encounter it every time we get onto a network- as soon as we connect to a router, to a website, send an email, make a skype/voip call etc.. In practice, the most used protocols for routing are based on well known standard graph distance finding algorithms such as Djikstra’s and Bellman-Ford, which itself is a testament to the longevity of these algorithms and to the power of graph algorithms, in general. If a node x gets a packet which started at node a and needs to end at node b, node x will refer to it’s routing table – a table which tells it which of its neighbours to send its packet to.
Often, a routing table will contain an entry for every node in the network telling where to forward a message addressed to that node. Now, this means the table can be really really huge depending on the size of the network. In practice, there are ways around this. One way, which makes for some nice theory is to do some preprocessing on the network e.g. build spanning trees, do DFS traversal, maybe some renaming and port changes, to reduce the size of the routing tables and the packet header. The crux of many of these schemes seems to be (the now seemingly simple) idea of interval routing (which by itself may not be compact) introduced by Santaro and Khatib in 1982. The idea may be summarised as follows:
- Starting from a particular node, do a Depth First Search (DFS) traversal and construct the corresponding DFS tree. Also, give each node a label that is that node’s DFS number (ID) (say, the time step at which that node was first encountered in the DFS traversal).
- Now, at each node, if you store the ‘largest’ ID of its subtree and the IDs of its children, you can now get intervals – which tell you which neighbour in the tree a node should send a packet addressed to another node. Hence, the name Interval Routing.
This spawned an active and productive field of research for the last few decades particularly in the effort of reducing both the size of the labels and tables while reducing the necessary tradeoff to be paid in terms of the distance. It is well known that if you reduce the space you use for routing, you cannot use the shortest paths and must pay in some measure by using a longer path (this measure is called stretch).
Looking at the above description, one will notice that developing these data structures seem to rely heavily on the initial DFS traversal. This implies that one may need to do a lot of recomputation if the network changes. In this spirit, there are some (but not many) works on `resilient’ compact routing i.e. compact routing that can handle changes to the network. Ours is one such attempt – in which we show how to do the well know Thorup-Zwick routing (over trees for now) in the self-healing model. In brief, the self-healing model is a responsive model of resilience where an adversary chooses a node/processor to take down or even insert (presumably to do the most damage in either case) and the neighbours of the attacked node/processor react by adding some edges/connections to the graph/network. We show how to both the compact routing and self-healing in low-memory (O(log^n) at most, where n is the number of nodes in the network).
I will defer the technical details of both self-healing and our compact self-healing routing to another post but needless to say, I am quite excited about continuing this work!