Yes sure. I have fixed the bug with the repeat stopping condition but
I have only tested pagerank on my small cluster. I still need to fix
the k-means clustering (it's a special case because you improve a
fixed number of points).
Leonidas
On Aug 30, 2012, at 9:02 AM, Edward J. Yoon wrote:
Shall we work together?
On Fri, Aug 24, 2012 at 9:01 PM, Leonidas Fegaras
<[email protected]> wrote:
Thank you very much for your interest and for testing my system.
It seems that my release was premature: It worked for some random
data but
didn't for some others. It's a minor logical error that I will try
to fix in
the next few days. The problem is with the stopping condition of
the repeat
expression that calculates the new pagerank from the old. It must
stop if
ALL peers reach the specified precision. This is done by having
those peers
that need to continue send a message to others to continue. It
seems that
now when all peers agree at the same time, the program works fine.
But if
one finishes sooner, instead of continuing the repeat loop, it runs
away to
the next BSP step that follows the repeat, then exits prematurely
and the
system hangs. The casting errors are due to the run-away peers
executing the
wrong BSP steps reading wrong messages. Queries without repeat
though are
OK.
By the way, I had a problem exchanging large amount of data during
sync (I
discussed this with Thomas). My solution was to to break a BSP
superstep
into multiple substeps so that each substep can handle a max number
of
messages. Of course my program has to collect all messages in a
vector in
memory. When the vector is too big, it is spilled in a local file.
This
moved the problem from the Hama side to my side and allowed me to
handle
larger data, especially in joins. I think this problem of
exchanging large
amount of data during a superstep is currently a weakness of Hama.
Leonidas
On 08/24/2012 04:15 AM, Thomas Jungblut wrote:
BTW, should we feature this on our website?
2012/8/24 Thomas Jungblut <[email protected]>
Hi Leonidas!
I have to admit that I have known what is going on (and had to keep
silent), but I have to say: Thank you very much!
This will help many people writing BSPs in a more easier way.
Of course this is not as fast as the native BSP code, Hive and
Pig suffer
from the same problems in MR.
But it gives people the opportunity to develop faster and get
their code
in production with just a minor time expense.
And I think, that we will help you gladly on improving the BSP
part of
your framework. At least I would do ;)
Thanks!
2012/8/24 Edward J. Yoon <[email protected]>
Here's my few test results on Oracle BDA (40G/s infiniband
network).
It seems slow than our PageRank example.
P.S., There are some errors so I couldn't test large-scale.
(java.lang.ClassCastException: hadoop.mrql.MR_int cannot be cast
to
hadoop.mrql.Inv and java.lang.Error: Cannot clear a non-
materialized
sequence ..., etc.)
== 100K nodes and 1M edges ==
*** Using 10 BSP tasks (out of a max 10). Each task will handle
about
2383611 bytes of input data.
Run time: 30.384 secs
*** Using 20 BSP tasks (out of a max 20). Each task will handle
about
1191805 bytes of input data.
Run time: 24.412 secs
On Fri, Aug 24, 2012 at 9:36 AM, Edward J. Yoon <[email protected]
>
wrote:
Wow, very interesting. I'm going to install and test on my large
cluster.
On Fri, Aug 24, 2012 at 4:41 AM, Leonidas Fegaras <[email protected]
>
wrote:
Dear Hama users,
I am pleased to announce that the MRQL query processing system
can now
evaluate SQL-like queries on a Hama cluster. MRQL is available
at:
http://lambda.uta.edu/mrql/
MRQL (the Map-Reduce Query Language) is an SQL-like query
language for
large-scale, distributed data analysis. MRQL is powerful
enough to
express most common data analysis tasks over many different
kinds of
raw data, including hierarchical data and nested collections,
such as
XML data. MRQL can run in two modes: in MR (Map-Reduce) mode
using
Apache Hadoop and in BSP (Bulk Synchronous Parallel) mode
using Apache
Hama. Both modes use Apache's HDFS to read and write their data.
Note that, the BSP mode is currently experimental (not fine-
tuned yet)
and lacks any fault-tolerance (if an error occurs, the entire
job must
be restarted). Due to our limited resources, MRQL has only
been tested
on a small cluster (7-nodes/28-cores). We compared the BSP
mode with
the MR mode by evaluating a pagerank query over a small graph
(100K
nodes, 1M edges) and found that BSP mode is about 4.5 times
faster
than the MR mode. Please let me know if you'd like to
contribute to
this project by testing MRQL on a larger cluster.
Best regards,
Leonidas Fegaras
University of Texas at Arlington
--
Best Regards, Edward J. Yoon
@eddieyoon
--
Best Regards, Edward J. Yoon
@eddieyoon
.
--
Best Regards, Edward J. Yoon
@eddieyoon
