Hi,
I am seeking some clarification about the networking models in use within
GridSim 4.2beta. That is the NET_FLOW and the NET_PACKET.
I have read (and experienced) that the FLOW model is faster to execute, and
scales better for larger files. However, I would expect that both of these
models should be equivalent in other respects. However, in the tests which I
have done I've found this is not the case.
A simple example is submitting a Gridlet with a input size, s, and timing
the ACK. ie, timing how long it takes to send s bytes. The FLOW model gives
me the answer I expected, the PACKET model was a factor of 10 too long.
Neither scaled as expected either (ie, doubling/tripling the data size did
not increase the ACK time by these factors. This was all done on a simple 1
router network with no traffic.
1MB file across Bottleneck Bandwidth : 1.0E7 bits/s
FLOW = 0.86s, PACKET = 8.71s
The second, and main issue I had was using the FLOW model for non-Gridlet
traffic the simulated tranfer time drops to be the propagation delay alone,
regardless of the data size. ie, if you Datagrid example 3 and run it with
the PACKET model the simulation time is 182578.99051989397. If you convert
all the routers and links to FlowRouters and FlowLinks and set the FLOW
model, the simulation time becomes 7203.992519893901. Essentially because
all the data transfers are done using the CTLG_REPLICA_DELIVERY (?) packet
type.
So, are my above observations accurate? And if so, should I just use the
PACKET model for datagrid/non-Gridlet simulations?
I am working on the assumption here that the scaling issues are close enough
to realistic for comparing scheduling algorithms, although my code which
estimates transfer times based on InfoPackets is obviously shot.
Many thanks
Nigel Sim
|
