Re: [Gems-users] where messages are actually moved?

[Liqun Cheng <liqun.cheng@xxxxxxxxx>]

What's the definition of switch latency, does it also include the
latency of interconnect (wires)? For example, considering a message
from L1a -> Switch 1 -> L1b. Switch latency is the latency from
Switch1's inport to its outport, or the latency from L1a's outbuffer to
L1b's inbuffer?

Thanks!
Legion

On 9/23/05, Bradford Beckmann <beckmann@xxxxxxxxxxx> wrote:

When you call MessageBuffer.enqueue(), (MessageBuffer.h, line 112) without
passing it a latency parameter, the latency defaults to one.  Therefore
the latency of messages enqueued by the PerfectSwitch is one.  Obviously a
switch latency of one is pretty optimistic, but that is why we call it a
"Perfect Switch".

Brad


On Fri, 23 Sep 2005, Liqun Cheng wrote:

> Hi there:
>
> I want to dynamically change the network latencies for some specific
> messages to model one normal network and one slow/power-optimized
> network (double the point-to-point latency).
>
> IMHO, all network latency changes are in the Throttle::wakeup, where
> the messages are actually moved from in_buffer to out_buffer
> m_out[vnet]->enqueue(m_in[vnet]->peekMsgPtr(), m_link_latency);
> m_in[vnet]->pop();
>
> However, I am puzzled about the relationship between Throttle::wakeup
> and PerfectSwitch::wakeup. In PerfectSwitch::wakeup, line 319,
> messages are also moved, although no latencies are added.
>
> Can somebody clarify this? thanks a lot!
> Legion
>