A Tuesday 07 December 2010 04:13:10 Philip Winston escrigué:
> Yes I think hyperslabs would be an essential tool for us.
> 
> Our hyperslabs would essentially be just single rows.  That worries
> me a little how it would perform but I should just try it and see. 
> We need to be able to extend the dataset so we need chunking just
> for that, if nothing else.
> 
> As per my other email I am worried maybe reading/writing single rows
> is not a good fit for HDF5?  But again I should really just
> experiment and see. Thanks.

My experience on this regard is that, when you want speed, there is 
little that can compete with mmap in terms of performance.  However, 
many times you may want to sacrifice extreme performance for more 
functionality.  In addition, memory mapping also has drawbacks, being 
the most important one the inability to map files that are larger than 
your available virtual memory, which renders this technology inadequate 
for many uses.

What many bindings for high-level languages are doing (specially in 
Python) is to treat datasets on-disk (available at low-level via HDF or 
NetCDF libraries) like if they were datasets in-memory.  That way, you 
are effectively dealing with on-disk data as if it was in-memory (this 
is what you are after, IIUC).  The OS filesystem cache is then in charge 
of caching as much data as possible in memory, so you get a behaviour 
that is very close in performance to a memory map approach.  Of course, 
you still have the HDF/NetCDF/whatever layer, which introduces some 
overhead, but this is largely compensated by other nice features, like 
on-the-flight compression (that may effectively accelerate I/O to disk) 
or practically unlimited dataset capacity (i.e. exceeding the virtual 
memory boundaries), among many others.

Regarding the comparison of binary formats (like HDF5) in comparison 
with relational databases, it is frequent the case that the former 
performs better than the later, specially if the interface is optimized.  

Hope this helps,

-- 
Francesc Alted

_______________________________________________
Hdf-forum is for HDF software users discussion.
[email protected]
http://mail.hdfgroup.org/mailman/listinfo/hdf-forum_hdfgroup.org

Reply via email to