home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.sys.isis
- Path: sparky!uunet!zaphod.mps.ohio-state.edu!rpi!batcomputer!cornell!honir!rcbc
- From: rcbc@honir.cs.cornell.edu (Robert Cooper)
- Subject: Re: Abcast performance
- In-Reply-To: hss@bigbird.bu.edu's message of 13 Aug 92 00:40:23 GMT
- Message-ID: <RCBC.92Aug12222023@honir.cs.cornell.edu>
- Sender: rcbc@cs.cornell.edu (Robert Cooper)
- Reply-To: rcbc@cs.cornell.edu
- Organization: Cornell Univ. CS Dept, Ithaca NY 14853
- References: <93520@bu.edu>
- Distribution: usa
- Date: Thu, 13 Aug 1992 03:20:23 GMT
- Lines: 98
-
- Himanshu,
-
- Your tests are explained by two factors:
-
- (1) The abcast algorithm used by the "bypass" protocols works like this;
- - sending member sends message to all members of group
- - each member puts this message in a queue
- - oldest member of group sends out a "sequencing" message to all
- members of the group. This ensure the total abcast ordering in the
- group, even with concurrent abcasts from different group members
- - on receiving the sequencing message, all members deliver the message
- to the application.
-
- In the case where the sending process is the oldest member the first step
- is skipped.
- This causes about a much faster response time between the oldest
- member of the group the the other members. In fact, smart Isis users
- often try to take advantage of this knowledge to speed up their programs.
-
- (2) By using abortreply you are probably confounding the results.
- When the sender gets the abortreply
- it stops listening for for further replies and returns immediately.
- When the oldest member of the group sends the abortreply, it does so
- without putting the message out on the wire. This internal software path
- is very fast. In the other-than-oldest-member case, the extra delays
- involved in the remote communication may well result in the nullreply
- getting to the sender before the sender has got its own abortreply back.
- The cost of handling the nullreply+abortreply might also explain some
- the differences. Its hard to say, but to remove this possibility,
- send the original message in "exclude sender mode", using abcast_l
- with the "x" option, and ask for ALL replies. This should achieve the
- effect you want of waiting for the remote member to reply.
-
- For the record, here was Himanshu's original posting:
-
- In article <93520@bu.edu> hss@bigbird.bu.edu (Himanshu Sinha) writes:
-
- Path: cornell!batcomputer!caen!uunet!olivea!bu.edu!bigbird.bu.edu
- From: hss@bigbird.bu.edu (Himanshu Sinha)
- Newsgroups: comp.sys.isis
- Subject: Abcast performance
- Summary: Abcasts sent by the first member of a group are 4-6 times faster than
- Message-ID: <93520@bu.edu>
- Date: 13 Aug 92 00:40:23 GMT
- Sender: news@bu.edu
- Distribution: usa
- Organization: Computer Science Department, Boston University, Boston, MA, USA
- Lines: 47
-
- Here is the gist of an experiment that I conducted.
-
- I have two processes P1 and P2 running identical programs running on different
- hosts. Each process waits (using isis_select) for the user to enter an integer,
- NumOfTImes, and makes NumOfTimes abcasts to the group "testgroup" of which both
- P1 and P2 are members. The abcast waits for 1 reply.
-
- Upon receiving a message, the receiver checks if the message has been
- sent by itself. If it is, it calls abortreply() otherwise it calls nullreply().
-
- I measure the response time for the NumOfTimes abcasts. P1 and P2 do not abcast
- at the same time.
-
- My observations were:
- 1) If P1 joined the group before P2 then its response time for a certain number
- of abcasts was about 4-6 times faster than the response time of P2 for the same
- number of abcasts.
-
- 2) If P2 joined the group before P1 then P2's response times were about 4-6
- times faster.
-
- Here are some numbers:
-
- When P1 joins before P2:
- Num of abcasts Response Time for P1 Respnse Time for P2
-
- 1 3653 usecs 19696 usecs
- 10 33606 usecs 141856 usecs
- 100 318263 usecs 1 secs 335790 usecs
- 1000 2 secs 811663 usecs 13 secs 606533 usecs
-
- When P2 joins before P1
- Num of abcasts Response Time for P1 Respnse Time for P2
- 1 12473 usecs 3182 usecs
- 10 134268 usecs 28624 usecs
- 100 1 secs 342105 usecs 268750 usecs
- 1000 13 secs 509179 usecs 2 secs 353236 usecs
-
-
- Question:
- How is this difference accounted for? I added a third process and a fourth
- process. The factor seemed to drop around 3 but I did not do enough reps to
- give any numbers.
-
-
- THanks
- Himanshu
- --
- --
-