Re: iSCSI: Towards Urgent Pointer Consensus

[Zack Best <zbest28@yahoo.com>]

> In cases 3) and 4) also, we do not need Urgent Pointer, because
> the bandwidth achieved will drop to a point (in response to packet
> drops), where you do not need a large amount of anonymous buffers.
>
> Venkat Rangan
> Rhapsody Networks Inc.
> www.rhapsodynetworks.com

 

I think this touches on an important point.  If one is running

a high bandwidth-delay product (lots of unacknowledged data in

the pipe at any given time), then the TCP congestion window must

be at least as large as the bandwidth-delay product, and the

packet loss probability must be sufficiently low to support

this size congestion window.  Using a simplified model, this

can be approximated as follows.

 

    bdp -> bandwidth-delay product

    p   -> maximum packet loss probability able to support bdp

    mss -> TCP maximum segment size 

 

    p = 2 * (mss / bdp)^2

 

Suppose instead of complicated out of order processing and TCP

fast retransmit recovery, we simply dropped all out of order

data and relied on the TCP timeout and retransmission to recover

the lost data?  Clearly the performance would suffer, but by

how much?

 

Again by using a simplified model, the time to recover is

as follows.

 

    rtt  -> round trip time

    rcvt -> recovery time

 

    rcvt = log2( bdp / mss) * rtt

 

The mean time between packet drops is:

 

    mtbd -> mean time between drops

    bw   -> average bandwidth which is equal to  bdp/rtt

 

    mtbd = mss / bw / p

         = mss / ( bdp/rtt ) / ( 2* (mss/bdp)^2 )

         = rtt * (bdp/mss) / 2

 

The net penalty ratio on performance is then (mtbd/rcvt).

Calculating this for a few points is interesting.  Assume

the mss is 1460 bytes, and rtt is 100ms.

 

 

bdp         bw        p            mtbd         rcvt        performance penalty

 

64KB      640KB/s   0.1%          2.2 sec.     0.55 sec.          25%

1MB       10MB/s    0.0004%        36 sec.     0.95 sec.           2.5%

10MB      100MB/s   0.000004%     360 sec.    1.3 sec.             0.36%

 

These numbers suggest that it is quite reasonable for a NIC card to

advertize an arbitrarily large window, and buffer up to 1MB of

out of order packets and drop all out of order packets beyond that.

It also underscores that achieving high bandwidth delay products

through a TCP connection (or any internet connection using the

TCP style congestion avoidance mechanism) requires extremely

low packet loss rates.

 

Zack Best,

GCL Systems, Inc.

---

Do You Yahoo!?
Yahoo! Shopping - 
Thousands of Stores. Millions of Products.