Conference noted::decps

2389.0. "Questions about Disk Queue Length" by CHOWDA::GLICKMAN (writing from Newport,RI) Wed Feb 12 1997 14:26

    We are looking at the queue length on a disk that has PMDF (third
    party mail product), TeamLinks and Hyperchannel (third party UNISYS to
    VAX transfer product) on it.  This is a VAXcluster (3 VAX 7740s and
    2 VAX 8820s).  The VAX 7740s are all purpose (including the servers
    for TeamLinks and Hyperchannel).  The VAX 8820s run PMDF and alot of
    printer queues.  The 8820s seem to be CPU intensive (200%) running PMDF
    internet mail transfer.  Note, again TeamLinks is not served from
    the 8820s.
    
    When we produced the PSPA Custom Graph including all the nodes for the
    queue length on that disk we find the stacked values reveal:
    
      VAX           0600 - 200 time period
              
     VAX 7740 (1)   under .5
     VAX 7740 (2)   under .5
     VAX 7740 (3)   mostly under .5, several occurences higer
    
     VAX 8820 (1)  mostly under 1, occassional blips to 3 - 4 and one blip 6
     VAX 8820 (2)  mostly under 1, occassional blips to 4 - 5 and 2 blipsover 10
    
    Note, this was only for one day but I want to understand some of the
    concepts relating to queue length on disks.
    
    If the two 8820s have the higher queue lengths for a given interval are
    IOs on the other VAXs affected.  For example, if a user is in Teamlinks
    and attempting to send a mail message (not necessarily an Internet mail
    message) during that interval would they see slower response because the
    8820s are doing Internet mail IOs to that disk. 
    
    Also note, that the graph shows the MIN as 0.02, MAX as 18.15 and
    Average as 0.99 for 0600 until 2000.
    
    So based on these numbers is the disk in bad shape in regards to queue
    length for particularly TeamLinks users on the 3 7740s.  Or do I
    need consider looking at this disk's queue lenght over a week or
    month to really understand how it is?  Do we tune for the blips or
    do we work with the Average?
    
    Appreciating any responses.
    

The queue length of one system may be effecting the IO perf of another 
system.  The item that best tells this is D_RESPONSE_TIME.  Graph this in a 
line format non-stacked and comper it to the same graph format of the queue 
length.  This will tell you if there is any significant increase in 
response time and if that follows the queue lengths of other systems.

Jilly