Mailing List Archive: Cisco: NSP

[nsp] latency vs throughput

Index | Next | Previous | View Threaded

pakucable at nettaxi

Sep 14, 2002, 5:51 AM

Post #1 of 4 (468 views)
Permalink

[nsp] latency vs throughput

which one is true ??

1) Latency increase ---> throughput(bandwidth usage) increase
2) Latency decrease ---> throughput(bandwidth usage) increase

ml at vayner

Sep 14, 2002, 9:19 AM

Post #2 of 4 (473 views)
Permalink

Re: [nsp] latency vs throughput [In reply to]

2 is correct

On Sat, 14 Sep 2002, pakucable wrote:

> which one is true ??
>
> 1) Latency increase ---> throughput(bandwidth usage) increase
> 2) Latency decrease ---> throughput(bandwidth usage) increase
>
> _______________________________________________
> cisco-nsp mailing list real_name)s [at] puck
> http://puck.nether.net/mailman/listinfo/cisco-nsp
> archive at http://puck.nether.net/pipermail/cisco-nsp/
>

gert at greenie

Sep 15, 2002, 12:08 PM

Post #3 of 4 (458 views)
Permalink

Re: [nsp] latency vs throughput [In reply to]

Hi,

On Sat, Sep 14, 2002 at 06:21:59PM +0530, pakucable wrote:
> which one is true ??
>
> 1) Latency increase ---> throughput(bandwidth usage) increase
> 2) Latency decrease ---> throughput(bandwidth usage) increase

As long as the window size of the L4 protocol in use is big enough,
the throughput should be independent of the latency.

If the window is too small, "2" is true(ish), but other factors might
apply as well.

gert

--
Gert Doering
Mobile communications ... right now writing from * back @ home *

ltd at cisco

Sep 15, 2002, 3:04 PM

Post #4 of 4 (451 views)
Permalink

Re: [nsp] latency vs throughput [In reply to]

At 06:21 PM 14/09/2002 +0530, pakucable wrote:
>which one is true ??
>
>1) Latency increase ---> throughput(bandwidth usage) increase
>2) Latency decrease ---> throughput(bandwidth usage) increase

it depends on the context.

in reality, Latency & Throughput are two completely different metrics
associated with "performance".
Latency isn't related to Throughput and vice-versa.

having said that, if the context is "related to TCP", then the most likely
correct answer is (2) --
TCP is a window-based protocol with maximum performance capped by the
"bandwidth * [rtt] delay" hitting the maximum number of bytes outstanding
on the wire allowed by the TCP window-size.

given a suitably-sized TCP window-size as well as enough 'go-fast' TCP
options (typically RFC 1323 including S/ACK & Fast Retransmission), most
real-world increases in latency can be made to result in minimal decrease
in throughput.

cheers,

lincoln.

Index | Next | Previous | View Threaded

Interested in having your list archived? Contact Gossamer Threads