Hello Stephen,
sorry for the very late reply.
On 28 Jun 2012, at 02:50, Fleischman, Stephen (ISS SCI  Plano TX) wrote:
> Hello Avery and all:
> I have a cluster of 10 twoprocessor/48 GB RAM servers, upon which we are conducting
Hadoop performance characterization tests. I plan to use the Giraph pagerank and simple shortest
path example tests as part of this exercise and would appreciate guidance on problem sizes
for both tests. I’m looking at paring down an obfuscated Twitter dataset and it would save
a lot of time if someone has some knowledge on roughly how the time and memory scales with
number of nodes in a graph.
>
I can provide some suggestions for the kind of algorithm and data which does currently surpass
the scalability of giraph.
While the limits to my knowledge of Giraph and Hadoop are probably also to blame for this,
please see the recent discussions on this list,
and on JIRA for other indications that the scalability of Giraph needs improvement:
* post by Yuanyuan Tian in the thread "wierd communication errors" on user@giraph.apache.org
* GIRAPH234 about GC overhead https://issues.apache.org/jira/browse/GIRAPH234?page=com.atlassian.jira.plugin.system.issuetabpanels:alltabpanel
If you want to stretch the limits of Giraph, then you need to try an algorithm which is conceptually
different from PageRank, and you need a big data set.
If you use an algorithm which has complex application logic (maybe even domain specific logic),
which needs to be embedded in the algorithm,
then the nodes need to have a lot of state. In addition, such algorithms probably send around
a lot of messages, and each of the messages might have a payload
which is more complex then one floating point number. In addition, it helps to have a graph
format, which requires strings on the edges and vertices.
The strings are required for the domain specific business logic which the graph algorithm
needs to follow.
Finally, imagine a data set which has a big loading time, and where one run of the algorithm
only provides results for one user.
The standard Hadoop paradigm is to throw away the graph after loading it.
So if you have 100s or 1000s of users, then you need a way to execute the algorithm multiple
times in parallel.
Again this will add a lot of state, as each of the vertices will need to hold one state object
for each user who has visited the vertex.
In my specific case, I had the following data and algorithm:
Data:
* an RDF graph with 10 million vertices and 40 million edges
I used my own import code to map the RDF graph to a undirected graph with a limit of one edge
between any two nodes (so it was not a multigraph)
* each vertex and each edge uses a string as an identity to represent a URI in the RDF graph
(required for the business logic in the algorithm)
Algorithm:
* spreading activation.
You can think of it as depth first search guided by domain specific logic.
A short introduction here: https://en.wikipedia.org/wiki/Spreading_activation
The wikipedia article only mentions using spreading activation on weighted graphs, however
I used it on graphs which have additional types on the edges.
The whole area of using the semantics of the edges to guide the algorithm is an active research
topic, so thats why I can't point you to a good article on that.
* parallel execution:
I need to run the algorithm once for every user in the system, however loading the data set
takes around 15 minutes alone.
So each node has an array of states, one for each user for which the algorithm has visited
a node.
I experimented with user numbers between 30 and 1000, anything more did not work for concurrent
execution of the algorithm.
Infrastructure:
* a single server with 24 Intel Xeon 2.4 GHz cpus and 96 GB of RAM
* Hadoop 1.0, pseudodistributed setup
* between 10 and 20 Giraph workers
A few weeks ago I stopped work on my Giraph based implementation, as Giraph ran out of memory
almost immediately after loading and initialising the data.
I made sure that the Giraph workers do not run out of memory, so it was probably due to IPC
and messaging.
The general discussion on the Giraph mailing list strongly indicates that I did hit the current
IPC scalability limits.
Currently I am working on a nonHadoop version of the algorithm which is not as scalable but
which is fast for *one* user. ( less then a second per user, but single threaded).
In addition, this new version allows me to better integrate with an existing ecosystem of
technologies (Semantic Web technologies) to which Hadoop and Giraph is currently completely
disconnected.
However, I will probably revisit Giraph at some time on the future.
If you want to look at the code or the data or any other asset which I have, then I will gladly
share that with you.
I would really like Giraph to reach the maturity required for this kind of algorithm.
However, I have the feeling that the current development focus is on clearcut numerical algorithms
such as pagerank.
