Browse Prior Art Database

Guideline for Flow Window Configuration in Data Redistribution

IP.com Disclosure Number: IPCOM000110067D
Original Publication Date: 1992-Oct-01
Included in the Prior Art Database: 2005-Mar-25
Document File: 3 page(s) / 117K

Publishing Venue

IBM

Related People

Chen, WT: AUTHOR [+2]

Abstract

Disclosed is a method for configuring a flow control window to improve performance in data redistribution when adding a node to a loosely coupled multi-node database system. In data redistribution operation, a newly added node has one logical link with each existing node that participates in the data redistribution operation. For each logical link, a window count is configured in the sending node and the receiving node, for communication effectiveness. Since the receiving node has to allocate window buffers for each logical link, the total buffer space the new node has to allocate for this data redistribution operation can grow rapidly as the number of existing nodes increases.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 52% of the total text.

Guideline for Flow Window Configuration in Data Redistribution

       Disclosed is a method for configuring a flow control
window to improve performance in data redistribution when adding a
node to a loosely coupled multi-node database system.  In data
redistribution operation, a newly added node has one logical link
with each existing node that participates in the data redistribution
operation.  For each logical link, a window count is configured in
the sending node and the receiving node, for communication
effectiveness.  Since the receiving node has to allocate window
buffers for each logical link, the total buffer space the new node
has to allocate for this data redistribution operation can grow
rapidly as the number of existing nodes increases.  To effectively
facilitate the data redistribution operation, a buffer allocation
scheme is proposed to determine the buffer space for each link
involved in the data redistribution process.

      As the number of messages coming from existing nodes is likely
to be very high, the receiving node is possibly the point of
contention.  With the window flow control mechanism considered, the
receiving node is a single server queue. It is assumed that the
messages arriving to the receiving node form a poisson process.
Also, the service time for each message is assumed to be a constant
value.  Based on these assumptions, the data redistribution process
can be considered as a M/D/1 queueing model.

      Considering the case of moving data from table T in node I to a
new node, the average arrival time of messages, denoted as Ri, is the
maximum of S and X, e.g., Ri = max {S,X}, where S is the average
time for preparing data in a message buffer and for routing the
message through the protocol stack; and X is the average time for
transmitting a message in the communication network.  X is usually
very small, especially when using a high-speed network like FDDI.

      The queueing time (denoted as Q) on the receiving node is the
sum of the waiting time and the service time. The average service
time (denoted as C) includes the message service time (the time for
receiving frames through protocol stack), the time for handling data
records in the received message, and the time for sending acknowledge
packet to the sending node.  As the sending node can quickly process
a received acknowledge packet and free the corresponding
communication buffer, the average time for this operation is short
enough to be neglected.

      Looking at the steady state of this M/D/1 queueing model, the
mean queueing time for each message becomes the same, regardless of
the source of the message.  When the queueing time is long and the
window count is small, it is possible that a sending node will be on
hold.  Suppose that Ki buffers are available at node I.  It is
obvious that t...