Re: Question about remote TFile usage

Constantin,

Reading from a local disk is of course faster (unless your local disk is slower than the remote disk). So, if you plan to read the entire file multiple times, you better copy it to a local disk if you have the network bandwidth to do it. If you access only small subsets of a file, remote access is better.

Rene Brun

Constantin Loizides wrote:
> Thanks, Rene,
>
> so the advantage of the "in-house" solutions
> are that optimized access is provided.
>
> For the other mentioned case I will have to
> test whether first copying to some local
> directory and then using TFile is advantageous
> compared to using the protocol provided by the
> TFile plugin. Right?
>
> Constantin
>
>
> Rene Brun wrote:
>> Constantin,
>>
>> When using versions of ROOT >= 5.14 instrumented with the TTreeCache,
>> you will get a very good access to remote ROOT files, including WANs
>> with high latencies.
>> When ROOT is used with xrootd (in particular the most recent
>> versions), you will even get better performance because in this case
>> we can use :
>> -clever asynchronous do read -aheads
>> -parallel streaming
>> With xrootd, in case of network problems, there is a redirector that
>> can automatically reconnect
>> to the same file via the same or different IP address, making the
>> system far more robust.
>>
>> Rene Brun
>>
>>
>> Constantin Loizides wrote:
>>> Hi all,
>>>
>>> I have a simple question:
>>> TFile access is typically and in particular in
>>> connection with trees very efficient.
>>> For example when reading only a branch out
>>> of a tree to first order only the data on
>>> disk that belong to the branch are read.
>>>
>>> Now my question is: How much of this
>>> efficient data access is kept when reading
>>> data remotely. For example comparing
>>> rootd, dcache and castor access?
>>> In case access to data is optimized
>>> such that only as much as data is transferred
>>> as needed, then I assume the connection to
>>> the server delivering the data has to
>>> be open at all times (via sockets?)
>>> Or will there be reconnects if needed?
>>>
>>> Thanks,
>>> Constantin
>>>
>>
>
Received on Thu Nov 01 2007 - 15:27:51 CET