Martin, you suggest:
If we do a fetchjoin(L,R) to a table R that does not fit in memory, and the operand L has more oids then the number of pages in R then we should also sort L on the tail oids first
The problem is that many MonetDB use fetchjoin as an order-aware operator and expect the output to be in left input order. This is then exploited in tuple reconstruction. It will certainly break pathfinder, for instance. Applications like Proximity would also break. In my thesis, I had suggested to create an accelerator that attaches a radix-clustered version to the left input. Fetchjoin would then use this radix-clustered version, and call decluster on the result, restoring the original order (using a cache and dick-friendly pattern). However, I have come to dislike the idea of low-level accelerators. I now think this is better handled on a higher layer, specifically the relational algebra layer. Peter
-----Original Message----- From: Martin Kersten [mailto:Martin.Kersten@cwi.nl] Sent: Friday, April 03, 2009 12:42 PM To: monetdb-developers@lists.sourceforge.net; Peter Boncz Subject: Re: [Monetdb-checkins] MonetDB/src/gdk gdk.mx, , 1.278, 1.279 gdk_bat.mx, , 1.213, 1.214 gdk_posix.mx, , 1.168, 1.169 gdk_relop.mx, , 1.161, 1.162 gdk_utils.mx, , 1.241, 1.242
Update of /cvsroot/monetdb/MonetDB/src/gdk In directory 23jxhf1.ch3.sourceforge.com:/tmp/cvs-serv31767/gdk
Modified Files: gdk.mx gdk_bat.mx gdk_posix.mx gdk_relop.mx gdk_utils.mx Log Message: various tweaks to the memory mangement
- tune the GDK_mmap_minsize to vary between full RAM size and 128MB (lowest when memory pressure is extreme)
- in fetchjoin, when the optimization of string heap copying is applied (instead of creating a new string heap and inserting into it), try to share the heaps in virtual memory (rather than copying it) This is possible with the logical view mechanism.
now fetchjoin, batcopy and remap use this VM heap sharing. If we do a fetchjoin(L,R) to a table R that does not fit in memory, and the operand L has more oids then the number of pages in R
Peter Boncz wrote: then we should also sort L on the tail oids first
This under the constraint that sorting L also cost some IO
If you are tweaking in that area anyway, such a patch is appreciated.
- disable vmalloc on Linux -- well actually on platforms that do not have posix_fadvise. A bit of a hack.
No virus found in this incoming message. Checked by AVG - www.avg.com Version: 8.0.238 / Virus Database: 270.11.39/2038 - Release Date: 04/02/09 19:07:00