Operation in non-steady workload.
In good days I row, in bad days I row harder
--
This post describes
the operation of the Architecture model when the nodes are working in a non-steady workload condition.
The situation where the
nodes’ workload can be considered non-steady is addressed in
Appendix “Nodes’ workload over time”.
Figure 3-1 of the
previous post "Requirements" is included here for convenience (see
Figure 7-1).
--------------------
In Figure 7-1, note that
the input node, the processing nodes set and the output nodes are
stages of a pipeline. The Throughput of each stage has to be at least
equal to the Throughput of the previous one. The Throughput of the
input node, processing node set and output node has to be at least
equal to the Throughput of the input data, input node and processing
node set, respectively.
 |
| Figure 7-1 |
When a node works under
non-steady workload condition, it experiences temporal work overloads
(work peaks). When a node experiences a work peak, its Latency grows.
From an external point of view, that node seems to become slower. If
the number of buffers for the transferences has not been calculated for the worst case,
when a destination node can’t absorb the output Throughput of its
source node, the faster will run the slower over with the result of
data loss. In cases like this, the data loss can be avoided by
incrementing the number of memory buffers.
Indeed, note that if the
work peaks are strong enough, the effect is similar to the one caused
by a traffic jam. The jam advances in the opposite direction to the
direction in which the cars advance. If for some reason, the output
node gets so overloaded that it can’t collect the data blocks from
the processing nodes, the lack of empty output buffers forces the
processing nodes to stop processing. If the situation persists, the
input node will stop distributing data blocks due to the lack of
empty input buffers in the processing nodes. Once it has exhausted
its own memory buffers, the input node will not be able to service
the input link and data loss becomes inevitable.
In the writing of this
article, The Who (Love Reign O'er Me, Quadrophenia) have collaborated
in an involuntary but decisive way.
--------------------
1. Picture:www.eitb.eus/es/.../video-un-dia-inolvidable-espectacular-bandera-la-concha-2017/
2. I want to thank Carol
G. for her revision of this text.
0 Comments