Which one has the lowest latency :

A - Cut-through
B - Fragment Free

Merci,

Soleil.

Report this post to a moderator

I would think Cut-through (Also wire-speed) has lowest latency, since it only reads the destination address before fowarding.
- Fragament-free (Modified cut-through) will wait for 64 kb window , before fowarding frames.
Any other opinion?

__________________
"No Work is done in vain!
The only place where Reward comes before
Work is in the Dictionary"

Report this post to a moderator

You are correct. Cut-through indeed does have the lowest latency for exactly the reason you described. Latency is a compromise with efficiency, however. Because you are transmitting the same amount of data over a higher number of frames with cut-through, it is less efficient as a whole because of more frame overhead, but at the same time it is a better solution for time-sensitive applications such as process control, video, and voice systems.

Strange, huh?

__________________
`·.¸¸.·´ Strike Attack `·.¸¸.·´
"If one strives towards a constant state of self-improvement, then the next day will always be better than the last. If the next day is better than the last, then tomorrow is destined to be a better day. If tomorrow is a better day, then one always has something to look forward to, one’s self has persevered, and true happiness becomes a little less difficult to attain." - Strike Attack

Report this post to a moderator

I believe (though I will hear back from many if I'm wrong) that because the chips and circuit design have gotten so fast, that store and forward is now deemed "wire-speed" though you can argue the technical inaccurcies of that statement, I believe it is now an accepted marketing term and the method used on most(if not all) newer switches these days.

My point being that it becomes moot about the speeds of fragment free and cut-through, because store and forward is what is used.

I'll standby for the wave of corrections

Yankee

Report this post to a moderator

i'll stand by everything the Yank says at least based on the grounds of him selling me on his theory .

as far as the test is concerned, know that, of your 3 methods of switching (cut-through, fragment-free, store and forward), cut-through does operate at wire speed and is considered fastest.

also make certain you understand some specifics of the other two types.

my Catalyst 5509 only does store and forward (if i remember correctly).

__________________
chunder

Report this post to a moderator

Thanks for all the infos,

Really appreciated.

Soleil.

Report this post to a moderator

I would agree with everything that Yankee says apart from one tiny and almost petty little point.

Yes latency is the lowest with Cut-Through switching and this is known as wire speed. Latency is highest (and variable) with Store and Forward. However, almost all new switches use Fragment Free as their default switching method. This is because almost everything that can go wrong with switching occurs within the first 64kb so by using Fragment Free the switches can incorporate a low latency with about an 99.5% certainty that the switching will be successful.

I hope that this helps.

Jamie

Report this post to a moderator

Hmmmmmm...I was sure the whole Catalyst series uses store and forward, but I could be wrong.

Yankee

Report this post to a moderator

Bare in mind that Cisco did not developed all the Catalist switches them selve, they kind of got little behind the switching revolution, so they bought, and bought, and bought. Then they put cisco labels on them. So the default switching mode will very with the level of service (type of Catalist). Many cisco implementations and that includes cisco routers, when it comes to Ethernet interfaces, will have Fast-switching mode by default. As you have mentioned above Fast-switching includes Cut-through and Fragment-free, from which Fragment-free is the default in use.

Fragment is a ethernet frame under 64 bytes, the minimum ethernet frame size. They are typicaly the leftover from collisions on a contentious segment such as non-switched LAN. A switched enviroment as one dictated by Fast-Ethernet, calls for full duplex, hub and spoke design, which eliminates collisions all toghether by providing virtual point to point links between 2 hosts that connect to the switch.

In regards of store and forward, many cisco products will switch automatically to store and forward from fast-forward mode when too many errors on the switched LAN are detected (that implies that some form of error checking is going on - fragment-free). I am not shure if it will reverse back to fast-switching, that I guess is configurable, or SNMP trap message is sent to the addministrator to alert him that the network is about to explode (there must be a reason why all of a suden too many errors are generated, malfunctioning interface/repeater/port). It is arguable where in the LAN store and forward would be of bennefit. Since the frame with an error will be dropped by the switch and no notification sent to the sender, thus the sending application will time out. Some applications are time sensetive and cannot tolorate such things - in such case it would be best to relay bad frame to destination, so the destination could notify sender immediatly, which in some cases would be faster then waiting for application timer to expire. Store and Forward is allways used in assynchronous switching, since the frame recieved from slow port has to be buffered before going out through a faster port and vice versa. Store and Forward is also a bennefit on highly congested networks, where it makes no sense to useup valuable bandwidth with dead frames that will be eventually dropped by destination, so this mode would be desirable in the backbone core.

When it comes to implementation of voice over IP, then fastest delivery is the most desirable. Since VoIP is real-time application, retransmition is out of the question here. Packetized voice will be typically caried in predictable, small sizes, that will allow for very small percentage of frames to be dropped, othervise the conversation will not be intelligable. Such job is beter suited to Layer 3 switch which will have the ability to look into IP header for presedence (priority bits that will allow some packets to be forwarded first before others). Again as the word implies looking up a IP header is a Store and Forward mode, but that only happens once in Fast-switching mode, once destination port is identified, all additional packets are switched imediatly to the destination with out any additional IP destination lookup. Switch will create route map to destination which will expire at some point if switch do not see any other traffic to destination.

Hope that will shine some light on that mysterious Store and Forward default setting.

GG, CCNA, CCAI.

Last edited by GilGrabber on 11-27-01 at 03:00 AM

Report this post to a moderator

Now you have changed the subject from layer two, to layer three when you bring in fast switching. This is apples and oranges....not the same subject.

Yankee

Report this post to a moderator