From kenmac Wed Nov 29 18:19:37 -0500 2000
From: "Kenneth M. Mackenzie" <kenmac@midway.cc.gatech.edu>
To: gte220x@prism, gte584f@prism, gt4947c@prism, gt8230b@prism, gte225u@prism, gte211q@prism, gte020r@prism, gte534w@prism, gte572q@prism, gte463q@prism, gt9182b@prism, gt0178f@prism, gt6645f@prism, gte823q@prism, gte148r@prism, gte909q@prism, gt5231a@prism, gte930q@prism, gte829m@prism, gte163i@prism, gt5193d@prism, gte182g@prism, gt7150b@prism, gte409s@prism, gte161h@prism, gt5809a@prism, gt6671b@prism, gt7116c@prism
CC: pp
Subject: clarifications on HW8
Reply-to: kenmac@cc.gatech.edu

Hi,

[note: paper for Friday if you didn't get it today:
http://www3-int.cc.gatech.edu/classes/AY2001/cs6290_fall/papers/codrescu99.pdf
http://www3-int.cc.gatech.edu/classes/AY2001/cs6290_fall/papers/codrescu99.ps
]

  Here are some clarifications about HW8 based on questions folks have
asked.  As usual, you can define anything you have to assume but here
are the assumptions I was making.

o  In Problem 1, the work per jacobi cell, w, is ambiguous.  I count
   one unit of "w" for each cell and ignore the number of
   multiplies/adds going on within a cell.

o  In Problem, parts D, E and F, ignore "edge" partitions in the
   array, i.e. compute the communication only for internal partitions
   that are bounded on all four sides, e.g, the partitions assigned to
   processors 5, 6, 9 and 10 in the figure on page 6.

o  In Problem 2, part A, the difference between synchronous and
   overlapped communication is ambiguous.

   I define synchronous communication as communication in which you
   always wait for round-trips.  For instance, to exchange a row of
   data with my neighbor, I send a message containing my row and then
   wait (doing nothing) for a return message containing the neighbor's
   row.  Only one message is in flight at a time.

   Overlapped communication assumes that a node does *not* wait for a
   reply when sending a message.  To send a message, I pay sender-side
   overhead plus transmit time but then after that I can immediately
   send another message (or receive a message).

o  Problem 2 mistakenly gives the computation per cell as three FADDs
   plus one FMUL.  There are really four FADDs.  I don't care which
   you assume but in the solutions I stuck with three FADDs rather
   than fixing it.

-- Ken