dear Experts,
i have one question related to DCACHE client side. we do have a DCACHE server in hand and we access our ROOT files in our analysis.
The file TTrees are chained and the analysis is performed by looping over the total number of entries of this TChain. fine.
i wonder how the mechanism is working indeed. DCACHE implements a set of environment variables which allow to tune the client side. We were asked to use these settings:
export DCACHE_RAHEAD=TRUE
export DCACHE_RA_BUFFER=32768
export DC_LOCAL_CACHE_BUFFER=1
export DC_LOCAL_CACHE_BLOCK_SIZE=131072
export DC_LOCAL_CACHE_MEMORY_PER_FILE=10000000
now my question. How does it really work. does the ROOT client DCAP/DCACHE interface at each GetEntry() of the TChain, ping the server and download the DCACHE_RA_BUFFER ? or does it download the file itself? i mean are the requests to the server event based, or are these based on some more clever approach?
thanks in advance, and sorry if i missunderstand.
some info:
5.26.00e_python2.6/i686-slc5-gcc43-opt
nm $ROOTSYS/lib/libDCache.so | grep dc_read
U dc_read
U dc_readdir
U dc_readv2
[quote]now my question. How does it really work. does the ROOT client DCAP/DCACHE interface at each GetEntry() of the TChain, ping the server and download the DCACHE_RA_BUFFER ? or does it download the file itself? i mean are the requests to the server event based, or are these based on some more clever approach?[/quote]It depends on whether you have turned on the TTreeCache or not. If you have not turned on the TTreeCache, ROOT will do small(ish) I/O request around 32K each time a branch needs a basket (a basket contains multiple event for a branch depending on the size of the data and the request size of the basket).
If you turn on the TTreeCache, then it will read much larger data chunk (the default is around 30Mb) in one I/O read. (it will read all the baskets for all the branches that are being read until it reaches around 30Mb).
Note that the whole mechanism have been improved significantly since v5.26 and we strongly recommend you upgrade to v5.30/02.