I am using cern 5.34.01 on windows. Most of my users read large, narrow TTrees over a LAN - often over 1E8 entries. These TTrees are rather narrow - typically 40-50 branches that are primitive data types. I stumbled across TTreeCacheUnzip and tried to use this feature - first followed the instructions in the 5.20 release notes (e.g. calling TTreeCache::SetParallelUnzip( TTreeCacheUnzip::kEnable)) but that caused a crash. I then found a forum post from 2003 that says to use gEnv->SetValue(“TFile.AsyncPrefetching”, 1); instead. I did that, nothing crashed, but I cannot tell if it is doing anything - I have not noticed any performance improvement.

My question is: Is this notion of having other cores unzip in parallel working/supported for windows? Are there any other things along these lines that I might try to speed reading up?

I have discovered that calling TTree::SetCacheSize() can speed things up by almost 3x if I use a buffer size of 20-30MB.

NB: back when I was using release 5.22, I was able to use xrootd running on a linux box close to the server… but sadly, TXNetFile is no long supported on windows…