Conference 7.286::fddi

What are you driving with and over what?

One should add that I made the claim that they shouldn't be able
to get anywhere close to 98 Mbit/sec. with 300s. Here are my
measurements, using "ttcp" and TCP/IP:

These are measurements on a two-station FDDI ring. One is  a 3000/300X (sand1)
the other a 3000/500 (kaputt), both with DEFTAs, running OSF/1 V3.0.

3000/300X -> 3000/500
ttcp -n 10000 -t -s -b 131072 -l 32768 kaputt-fddi
ttcp-t: 327680000 bytes in 41.54 real seconds = 7704.19 KB/sec +++

CPU usage on the sender is 99% system CPU, receiver is at around 53% CPU
load.

3000/500 -> 3000/300X
ttcp -n 10000 -t -s -b 131072 -l 32768 sand1-fddi 
ttcp-t: 327680000 bytes in 40.20 real seconds = 7960.43 KB/sec +++

This time the receiver is at 99% system CPU and the sender is at 48%
CPU load. 

Given that the CPU of the 300X is always the bottleneck with the slow
memory system and slow TurboChannel, I found it fair to extrapolate down
to the CPU speed of a 3000/300, which suggested to me

	150/175 * 7704 = 6603 Kbyte/sec.

which comes out to around 53 Mbit/sec.

re .2

   >      
   > 3000/300X -> 3000/500
   > ttcp -n 10000 -t -s -b 131072 -l 32768 kaputt-fddi
   > ttcp-t: 327680000 bytes in 41.54 real seconds = 7704.19 KB/sec +++
   >      

	327680000/41.54 = 7888.30 KB
	8*7888.3 = 63Mb
 
   > CPU usage on the sender is 99% system CPU, receiver is at around 53% CPU
   > load.
   >      
   > 3000/500 -> 3000/300X
   > ttcp -n 10000 -t -s -b 131072 -l 32768 sand1-fddi 
   > ttcp-t: 327680000 bytes in 40.20 real seconds = 7960.43 KB/sec +++
   >      
	
	327680000/40.20 = 8151.24 KB 
	8151.24*8 = 65 Mb

   > This time the receiver is at 99% system CPU and the sender is at 48%
   > CPU load. 
   >      
   > Given that the CPU of the 300X is always the bottleneck with the slow
   > memory system and slow TurboChannel, I found it fair to extrapolate down
   > to the CPU speed of a 3000/300, which suggested to me
   >      
   >      	150/175 * 7704 = 6603 Kbyte/sec.
   >      
   > which comes out to around 53 Mbit/sec.
   > 

	Why extrapolate down for this?

Bill 

    > Why extrapolate down for this ?
    
    The situation here is that someone seems to have promised the customer
    near 100 Mbit/sec. TCP/IP throughput on their 3000/300. I wanted to
    show that even a 3000/300X doesn't get anywhere near that (more like
    65 Mbit/sec.) and that it was reasonable to expect less than the
    3000/300X throughput, possibly scaling as CPU powers (even if that
    weren't true, the 300X numbers should be an upper bound).

	I got the following performance data from the Alpha OSF group at ZKO.
	They ran the 3000-300X data over the last two days.
	
	Both the 300X and the 500 have the same clock rate (150MHz). 
	The difference is the external cache (B-cache). The 500
	has 512 KB and the 300X has 256 KB. The 300X also has a smaller 
	memory Bus Width (64 bits vs 256 bits), Memory Bus Speed 
	(66 MB/S vs 114 MB/S) and CPU Data Bus Width (64 bits vs 128 bits) 
	
	Basically, they are two different machines and with a better 
	system (500) we can push up to the FDDI wire limit of very
	close to 100 Mb (at 32 Kb message size the bandwidth is 
	12.112*8 = 96.89 Mb/S...see table below).



    +-------------------------------+------------------------+
    | Pair of 3000-300X's with DEFTA| Pair of 3000-500's with|
    | FDDI Adapters,                | DEFTA Adapters,        |
    | OSF V3.2-5, 128 kB Window Size| OSF V3.0, 128kB Windows|
    +-------+-------+-------+-------+--------+-------+-------+
    |Message|  TCP  | TCP   | TCP   |  TCP   | TCP   | TCP   |
    |  Size | Bandw |  SX   | RX    | Bandw  |  SX   | RX    |
    |       |  MB/s | CPU % | CPU % |  MB/s  | CPU % | CPU % |
    +-------+-------+-------+-------+--------+-------+-------+
    |    1  | 0.046 | 98.63 |  0.00 |  0.046 | 99.54 |  0.46 |
    |   64  | 1.661 | 76.28 |  7.19 |  1.660 | 68.35 |  7.08 |
    |   108 | 1.911 | 63.04 | 13.86 |  2.520 | 66.74 |  8.94 |
    |   112 | 2.254 | 68.76 | 10.67 |  2.552 | 62.72 |  8.81 |
    |   224 | 2.870 | 70.16 | 19.57 |  4.258 | 70.13 | 13.13 |
    |   512 | 4.623 | 76.50 | 24.30 |  6.528 | 72.18 | 21.45 |
    |  1024 | 5.261 | 81.59 | 32.21 |  8.619 | 78.13 | 24.98 |
    |  1460 | 5.712 | 86.15 | 29.56 |  8.724 | 77.63 | 26.89 |
    |  2048 | 6.020 | 84.73 | 37.39 | 11.072 | 89.23 | 37.64 |
    |  3072 | 6.520 | 83.32 | 46.27 | 11.378 | 75.37 | 45.86 |
    |  4096 | 6.626 | 81.61 | 49.94 | 12.127 | 76.32 | 47.75 |
    |  5120 | 6.740 | 83.74 | 46.82 | 11.672 | 73.78 | 45.69 |
    |  6144 | 6.810 | 84.64 | 45.24 | 11.798 | 70.90 | 43.82 |
    |  7168 | 7.105 | 85.64 | 44.94 | 11.804 | 69.39 | 46.28 |
    |  8192 | 7.314 | 84.72 | 45.73 | 12.106 | 70.46 | 48.64 |
    |  9216 | 6.964 | 87.13 | 43.66 | 11.877 | 71.50 | 46.52 |
    | 10240 | 6.992 | 87.91 | 46.55 | 11.933 | 70.04 | 45.38 |
    | 11264 | 7.188 | 87.08 | 47.91 | 11.915 | 69.82 | 47.69 |
    | 12288 | 7.271 | 84.93 | 51.31 | 11.967 | 69.26 | 49.34 |
    | 13312 | 7.388 | 86.52 | 50.61 | 11.944 | 67.70 | 47.64 |
    | 14336 | 7.406 | 87.03 | 52.26 | 11.988 | 67.46 | 46.32 |
    | 15360 | 7.416 | 87.58 | 50.83 | 11.983 | 67.25 | 47.54 |
    | 16384 | 7.538 | 88.39 | 51.99 | 12.101 | 67.29 | 46.20 |
    | 17408 | 7.458 | 90.23 | 51.40 | 11.979 | 67.47 | 49.09 |
    | 18432 | 7.397 | 89.39 | 50.80 | 12.026 | 66.90 | 45.00 |
    | 32768 | 7.748 | 91.80 | 54.69 | 12.066 | 66.21 | 46.93 |
    | 65536 | 7.295 | 93.24 | 50.79 | 12.112 | 67.09 | 47.10 |
    +-------+-------+-------+-------+--------+-------+-------+



    

     Correction to .5 -
    
            The 300X has 175 Mhz clock rate and not 150 (300 has 150 MHz).