Re: Overlapping put/get events

From: Dan Bonachea (bonachea_at_cs_dot_berkeley_dot_edu)
Date: Tue May 05 2009 - 18:36:59 PDT

  • Next message: Dan Bonachea: "Re: About barriers"
    At 06:42 AM 4/29/2009, Nikita Andreev wrote:
    >Hi, guys.
    >
    >The issue with filename, colnum and linenum is resolved. I just installed upc 
    >2.8.0 and alt l started to work. Anyway thanks for the help.
    >
    >I have some other questions.
    >
    >1. Is it possible in UPC that two put/get operations overlap? I need to match 
    >corresponding start/end events in trace file. Could I face situation when I 
    >have such trace:
    >UPC_PUT START
    >UPC_PUT START
    >UPC_PUT END
    >UPC_PUT END? And I don't know which start match which end.
    
    The PUT/GET events from a single thread should not overlap (those are the 
    "blocking" put/get events), but the NB (non-blocking) variants of those events 
    generated by the optimizer and BUPC library extensions can and will. Those 
    events include handle information specifically to make this matching possible.
    
    If you're interested in writing a UPC performance analysis tool, you should 
    really use the GASP interface directly, rather than relying on the minimal 
    trace output of the "toy" GASP dump tool. You'll miss alot of information and 
    pay unnecessary profiling performance overheads that way. For more info, see 
    http://gasp.hcs.ufl.edu/
    
    If you're just looking to analyze the performance of UPC apps, I'd highly 
    recommend the PPW GASP tool available at http://ppw.hcs.ufl.edu/
    
    >
    >2. There is a GASP_UPC_FORALL event in GASP 1.5 specification. But in 
    >gasp_upc.h header distributed with Berkeley UPC this event is not declared. 
    >Is it implemented?
    
    BUPC does not currently generate the GASP_UPC_FORALL event (all events are 
    "optional" by the GASP spec, and we focused our implementation on those most 
    central to parallel performance analysis, namely communication and 
    synchronization - line numbers are generally sufficient to determine 
    upc_forall nesting given the source file, if this matters).
    
    -D
    

  • Next message: Dan Bonachea: "Re: About barriers"