# Graph Algorithms and Software Prefetching

ModelingPredictive Analyticsposted by Daniel Lemire May 31, 2018

A lot of data in the real world can be represented as graphs: you have nodes connected through edges. For example,...

A lot of data in the real world can be represented as graphs: you have nodes connected through edges. For example, you are a node in a graph where friendships are edges.

I recently met with professor Semih Salihoglu, an expert in graph algorithms and databases. We discussed fun problem like how one can find the shortest path between two nodes in a very large graph.

[Related Article: Greater Speed in Memory-Bound Graph Algorithms with Just Straight C Code]

Semih argued something to the effect that, often, the best you can do is a breadth-first search. That sounds scary and fancy… but it is actually super simple. Start from a given node. This node is at level 0. Then visit all its neighbors (by iterating through its edges). These nodes are at level 1. Then visit all the nodes connected to the nodes at level 1 (excluding your starting node), these are at level 2. And so forth. The magic is that you will end up visiting once (and exactly once) all nodes that can be reached from your starting node.

With this approach, you can find the distance between any two nodes. Just keep exploring, starting from one of the two nodes, until you encounter the other node.

But what happens if the graph is very large? These nodes you are visiting are all over your memory. This means that each node access is a potential cache fault. Our processors are super fast, and they have super fast memory close to them (cache memory), but your main memory (RAM) is comparatively much slower.

Thus, when processing a large graph, you are likely memory-bound… meaning that much of the time is spent waiting for memory access. It is worse than it appears because memory access is a shared resource in multicore processors, which means that you cannot make this problem go away cheaply by buying a processor with many more cores.

Can we quantify this?

I built a large random graph made of 10 million nodes where each node has 16 random neighbors.

I pick a node at random, seek another node far away, and then I measure the time it takes to do the breadth-first search from one to the other. On a per-edge basis, it takes 23 cycles. Don’t worry, things get much worse if the graph gets larger, but let us reflect on the fact that 23 cycles to merely look at the node identifier, check if it has been visited and if not, add it to our list… is a lot. Not counting memory accesses, we should be able to do this work in 5 cycles or less.

Can we do better than 23 cycles?

What if, right before you start processing the neighbors of one node, you told your processor to go fetch the neighbors of the next node? I have a recent post on this very topic: Is software prefetching (__builtin_prefetch) useful for performance?

In that post, I explained that Intel processors have prefetching instructions that the software can call. I also recommended to avoid them.

So what happens if I add a prefetch instruction to my graph code? I go down to 16 cycles… saving a third of the running time.

 naive breadth-first search 23 cycles per edge visited prefetched breadth-first search 16 cycles per edge visited

My result would seem to invalidate my recommendation to avoid software prefetches. But keep in mind that my implementation is naive and limited, thrown together in a couple of hours. It is a proof of concept. What it demonstrates is that even if you are limited by memory accesses, there are still software choices you can make to help you.

I would only change my recommendation against software prefetches if we definitively could not rewrite the code differently to get the same benefits. I think we can write more clever code.

There are many problems with software prefetches. In some cases, as is the case here, it is better than nothing… But it is still a fragile hack. It helps in my particular case, but change the parameters of the graph, and things might go to hell. Update your processor and things could go to hell. And there is no way to know whether the exact way I did it is close to optimal… it works well in my case, but it might require much tuning in other instances.

So how can we write the code better? I am not certain yet.

Original Source

Stay up to date on all things data science here.

## Daniel Lemire

Daniel Lemire is a full professor in computer science at the University of Quebec (TELUQ). His research is focused on data indexing techniques. For example, he worked on bitmap indexes, column-oriented databases and integer compression. He is also interested in database design and probabilistic algorithms (e.g., universal hashing). His work on bitmap indexes is used by companies such as eBay, LinkedIn, Facebook and Netflix in their data warehousing, within big-data platforms such as Apache Hive, Druid, Apache Spark, Netflix Atlas, LinkedIn Pinot and Apache Kylin. The version control system Git is also accelerated by the same compressed bitmaps. Some of his techniques were adopted by Apache Lucene, the search engine behind sites such as Wikipedia or platforms such as Solr and Elastic. One of his hashing techniques has been adopted by Google TensorFlow. His Slope One recommender algorithm is a standard reference in the field of recommender systems. He is a beneficiary of the Google Open Source Peer Bonus Program. He has written over 50 peer-reviewed publications, including more than 30 journal articles. He has held competitive research grants for the last 15 years. He serves on the program committees of leading computer science conferences (e.g., ACM CIKM, WWW, ACM WSDM, ACM SIGIR, ACM RecSys).

1