linearregression / graphchi Goto Github PK

If GraphChi uses multiple disks, also vertex data needs to be read in a striped 
fashion.

Seems gzip or bzip2 can compress even a pagerank edge file very significantly

If file was adjacency, but try to read with edgelist format -> segfault.

All current example apps initialize their vertex data in the update function, 
on first iteration. For some applications, such as common in belief 
propagation, vertices have predefined values. This is easy to accomplish in 
GraphChi: simply write the [file-prefix].vout file before running the 
application. This file is just a flat binary array.

However, it would be useful to have a tool to parse a text file into the vertex 
data file, at least to show an example.

LPAm is a more advanced community detection algorithm based on label 
propagation. It requires a community count to be recomputed on every iteration. 
This should be relatively easy with the aggregators.

Requested by two users..

In some cases we assume that vertex-ids can be converted to 32-bit signed 
integers. At least in the sliding-shards jump-index.

Fix and test with 4B vertices!

What is the expected output? What do you see instead?
1.the output should tell us which community the node is belong to ,not only 
community-ids and their sizes,and it is hard to identify which nodes the 
community includes,which make it difficult to analysis the feature of the 
community due to lake of information about the nodes in it.

2.moreover,the output includes some other files (as attached below,file 0 to 
15,here only 0 and 3 for example)which cannot read. what do they really mean?

What version of the product are you using? On what operating system?
graphchi on ubuntu 12.0.4

It should be easy to write code to aggregate vertex values efficiently.

If dynamic scheduling is disabled, engine should use a mock-scheduler instead 
of having a null scheduler. 

- Add "selective_scheduling" to graphchi_context

Html-based admin is nice.

Requested by a few users..

do you have plan to develop a version for windows?

What is the expected output? What do you see instead?
1.the output should tell us which community the node is belong to ,not only 
community-ids and their sizes,and it is hard to identify which nodes the 
community includes,which make it difficult to analysis the feature of the 
community due to lake of information about the nodes in it.

2.moreover,the output includes some other files (as attached below,file 0 to 
15,here only 0 and 3 for example)which cannot read. what do they really mean?

What version of the product are you using? On what operating system?
graphchi on ubuntu 12.0.4

What steps will reproduce the problem?
1. Write a program that uses dynamic vertex data.
2. Run it on a big graph with many iterations (depends on the size of the graph 
and the available memory).
3. Observe the segfault that is caused by the memory leak.

What is the expected output? What do you see instead?
Our code ran on a graph with about 5000 vertices. We have discovered two memory 
leaks:
- at dynamicblock.hpp:106, which resulted in a loss of 37MB (valgrind.1)
- at graphchi_engine.hpp:992, which leaked about 230kB (valgrind.2)

What version of the product are you using? On what operating system?
Latest trunk.

Please provide any additional information below.
I have attached our fixes (patch).

Similar to how ALS works. In the beginning of the app, check if shards have 
been created, and if not, create them. 

Port GaBP code to graphchi linear solvers toolkit. Requested by Ori Shental.

Looking at this code snippet from example_apps/pagerank.cpp

-- -- >8 -- -- >8 -- -- >8 -- -- >8 -- -- >8 -- -- >8 -- -- >8 
#define RANDOMRESETPROB 0.15

// ...

struct PagerankProgram :
  public GraphChiProgram<VertexDataType,
                         EdgeDataType> {

  // ...
  void update(graphchi_vertex<VertexDataType, EdgeDataType> &v,
              graphchi_context &ginfo) {
    float sum=0;
    // ...
    for(int i=0; i < v.num_inedges(); i++) {
      float val = v.inedge(i)->get_data();
      sum += val;                    
    }

    /* Compute my pagerank */
    float pagerank = RANDOMRESETPROB + (1 - RANDOMRESETPROB) * sum;
    // ...
  }
}
-- -- >8 -- -- >8 -- -- >8 -- -- >8 -- -- >8 -- -- >8 -- -- >8 

looks like that the current pagerank implementation doesn't
allow you to set the "personalisation vector" to anything different
than a uniform probability vector. I mean, if the pagerank equation
in matrix form is

p = (1-c) * A * p + c * V

where:

p is the pagerank vector, N compontents (N is the size of the web)
c is the probability of jumping to a random page no matter the outlinks
  from current location
A is the transition matrix, N-by-N, if you see a random walk on the web
  as a Markov chain
V is a N-vector, where V_i is the probability of random-jumping to page i

(side note: I am not normalizing by N, i.e. all probabilities sum up to N
and not to 1)

well, given all of this, in the current implementation of GraphChi pagerank
V is a uniform probability vectory = [1, 1, 1, ..., 1].
A jump to every page is equally likely to happen, no matter the page.

After this wall of text I come to my point:

could a non trivial "personalisation vector" be implemented?
I'd like to be able to set V myself.

Is this in the priorities of the GraphChi team?

Cheers,

... which builds all example apps.

Make smoketest for the dynamic graph

We should convert static allocation to dynamic using command line argument 
specification of feature width

Currently deleted edges are not reflected until the window has passed.

What steps will reproduce the problem?
1. bin/exampleapps/pagerank file pagerank2.txt
2.
3.

What is the expected output? What do you see instead?
top vertices

What version of the product are you using? On what operating system?
Using 0.1b of graphchi on Mac OSX 10.5.6

Please provide any additional information below.

Currently it is required that all edge values are of same type and size. For 
some algorithms, it might be that some edges need more data than others. It is 
fairly easy to accomplish this by writing a size-word and perhaps a type-byte 
before each edge data value. However, this means that edge data must be loaded 
synchronously with adjacency data as we could not just compute the pointers.

How to implement different types or variable size containers cleanly in the C++ 
code is a question.

What steps will reproduce the problem?
bin/example_apps/pagerank file GRAPH-NAME

What is the expected output? What do you see instead?


What version of the product are you using? On what operating system?
linux

Please provide any additional information below.

my graph:
1   2   
1   3   
1   4   
2   4   
3   4   
3   5   
4   5   
5   4


top 20 vertices:

1. 5    0.604059
2. 4    0.437937
3. 2    0.1925
4. 3    0.1925
5. 0    0.15
6. 1    0.15
7. 6    0.15

If user has plenty of memory available, it is good idea to pin some shards to 
memory, so they need not to be reprocessed from disk.

I think we can avoid initialization on first iteration by being smarter when 
reading the edges.

I tried to build the sharder_basic program with the latest download and it 
failed with this error:

g++ -g -O3 -I/usr/local/include/ -I./src/  -fopenmp -Wall -Wno-strict-aliasing 
src/preprocessing/sharder_basic.cpp -o bin/sharder_basic
In file included from ./src/preprocessing/conversions.hpp:36,
                 from src/preprocessing/sharder_basic.cpp:35:
./src/preprocessing/sharder.hpp: In member function ‘virtual void 
graphchi::sharder<EdgeDataType>::write_shards()’:
./src/preprocessing/sharder.hpp:465: error: ‘degree’ was not declared in 
this scope
./src/preprocessing/sharder.hpp:467: error: ‘degrees’ was not declared in 
this scope
./src/preprocessing/sharder.hpp:469: error: expected primary-expression before 
‘)’ token
./src/preprocessing/sharder.hpp:469: error: expected `;' before ‘calloc’
make: *** [sharder_basic] Error 1

Including the degree_data header seemed to fix this and I was able to build and 
run basic_sharder after this

#include "engine/auxdata/degree_data.hpp"

Now if in-edges are not properly commited, the test can still succeed.

Create a wrapper/API to run Graphlab v2.1 vertex programs.

Assume that vertices are stored in memory, not replicated to edges. This will 
enable most graphlab apps to run without modification, although the memory 
consumption will be higher than on normal GraphChi applications.

What steps will reproduce the problem?

1. using sparse matrix available at
https://raw.github.com/gist/3422269/e0a585749a3a9ae84003fc368e80b69398a30d22/gis
tfile1.txt

2. ./bin/als training lsa_test.mm

What is the expected output? What do you see instead?

expected decomposition, instead got assertion failure.
for full output see https://gist.github.com/3422275

What version of the product are you using? On what operating system?

setup as described 
http://bickson.blogspot.co.at/2012/08/collaborative-filtering-with-graphchi.html
?m=1