[Assorted-commits] SF.net SVN: assorted: [415] hash-join/trunk

[<yan...@us...>]

Revision: 415
          http://assorted.svn.sourceforge.net/assorted/?rev=415&view=rev
Author:   yangzhang
Date:     2008-02-14 17:40:15 -0800 (Thu, 14 Feb 2008)

Log Message:
-----------
updated analysis, readme, doc publishing

Modified Paths:
--------------
    hash-join/trunk/README
    hash-join/trunk/doc/Makefile
    hash-join/trunk/doc/analysis.txt

Modified: hash-join/trunk/README
===================================================================
--- hash-join/trunk/README	2008-02-14 20:33:35 UTC (rev 414)
+++ hash-join/trunk/README	2008-02-15 01:40:15 UTC (rev 415)
@@ -29,6 +29,9 @@
   there is a match, then emit the resulting joined tuple (movie title, movie
   release year, actress name).
 
+Results
+-------
+
 Here are some [results].
 
 Requirements
@@ -75,7 +78,7 @@
 this dataset and to observe the resulting distributions.
 
 [C++ Commons]: http://assorted.sf.net/cpp-commons/
-[HashDist]: http://assorted.sf.net/
+[HashDist]: http://assorted.svn.sourceforge.net/viewvc/assorted/hash-dist/trunk/
 [Multiprocessor Hash-Based Join Algorithms]: http://citeseer.ist.psu.edu/50143.html
 [Scala Commons]: http://assorted.sf.net/scala-commons/
 [g++]: http://gcc.gnu.org/

Modified: hash-join/trunk/doc/Makefile
===================================================================
--- hash-join/trunk/doc/Makefile	2008-02-14 20:33:35 UTC (rev 414)
+++ hash-join/trunk/doc/Makefile	2008-02-15 01:40:15 UTC (rev 415)
@@ -1,6 +1,7 @@
-PROJECT := hash-join
-WEBDIR  := assorted/htdocs/$(PROJECT)
-PANDOC   = pandoc -s -S --tab-stop=2 -c ../main.css -o $@ $^
+PROJECT  := hash-join
+WEBDIR   := assorted/htdocs/$(PROJECT)
+HTMLFRAG := ../../../assorted-site/trunk
+PANDOC    = pandoc -s -S --tab-stop=2 -c ../main.css -H $(HTMLFRAG)/header.html -A $(HTMLFRAG)/google-footer.html -o $@ $^
 
 all: index.html analysis.html
 
@@ -14,10 +15,10 @@
 	ssh shell-sf mkdir -p $(WEBDIR)/
 	scp $^ shell-sf:$(WEBDIR)/
 
-publish-data: times.pdf speedups.pdf
+publish-data: ../tools/data/*.pdf
 	scp $^ shell-sf:$(WEBDIR)/
 
 clean:
 	rm -f index.html analysis.html
 
-.PHONY: clean publish
+.PHONY: clean publish publish-data

Modified: hash-join/trunk/doc/analysis.txt
===================================================================
--- hash-join/trunk/doc/analysis.txt	2008-02-14 20:33:35 UTC (rev 414)
+++ hash-join/trunk/doc/analysis.txt	2008-02-15 01:40:15 UTC (rev 415)
@@ -1,4 +1,4 @@
-% Hash-Join Benchmarks
+% Hash-Join Analysis
 % Yang Zhang
 
 Here are the graphs from the latest experiments and implementation:
@@ -9,7 +9,7 @@
 This implementation was originally not scalable in the hashtable-building
 stage, which performed frequent allocations. The hashtable is stock from the
 SGI/libstdc++ implementation. I removed this bottleneck by providing a custom
-allocator that allocated from a non-freeing local memory arena.
+allocator that allocates from a non-freeing local memory arena.
 
 Profiling reveals that most of the time is spent in the hash functions and the
 function that performs the memcpy during hash-partitioning. `actdb::partition1`
@@ -27,11 +27,14 @@
   ...
 
 Now the hashtable construction phase is the most scalable part of the
-algorithm. The remaining bottlenecks appear to be due to the memory stalls.
+algorithm (despite its random access nature). The remaining bottlenecks appear
+to be due to memory stalls, but these are mostly masked by hardware
+prefetching.
 
-The program does not scale much beyond the 16 threads, though performance does
-improve slightly. This is due to the contention for cache capacity among
-multiple hardware threads per core.
+The program does not scale much beyond 16 threads, though performance does
+improve slightly. The inability to scale beyond 16 is most likely due to the
+contention for cache capacity among multiple hardware threads per core.
 
-This implementation is straightforward, with no fanciness in terms of custom
-scheduling and control over allocation, leaving many things up to the OS.
+I've tried to keep the implementation simple, with no fanciness in terms of
+custom task scheduling or control over allocation, leaving many things up to
+the OS.


This was sent by the SourceForge.net collaborative development platform, the world's largest Open Source development site.