On Tue, 19 Nov 2002 15:24:51 -0500
Jan Harkes <jaharkes_at_cs.cmu.edu> wrote:

> On Tue, Nov 19, 2002 at 01:49:46PM +0100, Marcel Pol wrote:
> > It works fine with a size of the datapartition of 22M, 44M, 90M, 130M and
> > 200M, but it breaks when I select the 315M. The log size was 4M. 
> > The real log partition is 50M, and the data partition is 350M in cfdisk.
> > The real partition of /vicepa is 11Gb, and the /usr/coda has that size
> > too.
> > 
> > This is the error:
> > ?  ERROR: rds_zap_heap RVM_EINTERNAL.
> 
> It depends on the amount of real and swap memory your machine has. As
> you are using partitions instead of files, we cannot use a private mmap,
> and allocate a block of memory that is as big as the RVM data partition.

Thanks for the reply.
The machine has only 96Mb of ram, and 170Mb of swapspace, on another disk.
>From your explanation below I gather that it might be a better option to use
files instead of partitions then. 


> The complete file is then read into this allocated chunk of memory. When
> your available amount of real memory is significantly less than 315MB,
> the machine will have to swap a lot in order to do this. It reads a
> block from disk, then faults in a page, which is then written back to
> swap. If the RVM data and swap partitions are on the same disk this
> involves a lot of seeks.
> 
> If you were to use a file for RVM data and set the map_private flag in
> the server.conf file loading RVM is really quick because we only have to
> set up a private mmapped area. Data is paged in as needed, and only
> dirtied pages end up in the swap space. Ofcourse over time all pages
> would get dirtied and you still end up using as much swap as in the
> other case, but the hit is spread over the lifetime of the server and
> as a result the disk doesn't have to trash as much.



--
Marcel Pol

Linux 2.4.19-18mdk-ringworld.1, up 8 days, 23:41
Registered User #163523