| T.R | Title | User | Personal
Name | Date | Lines |
|---|
| 332.1 | Checksum? | STKHLM::TORGNY | | Tue Aug 27 1991 13:56 | 8 |
| Hi,
Check if SUN is using checksum, rumor is that SUN's default is off
and ours on. On UDP disabling checksum gives a 75% boost. See
Digital Technical Journal vol 3 num2 spring 1991 for FDDI performance.
Regards,
Torgny...
|
| 332.2 | from another conference | OZROCK::FARAGO | ncl-dns-cml-cmip...aargh!! | Wed Aug 28 1991 20:18 | 22 |
| From DNU-PHASEV-IFT notesfile
================================================================================
Note 222.6 What about FDDI 6 of 7
STKHLM::TORGNY 15 lines 12-JUN-1991 12:08
-< TP4 faster than NSP >-
--------------------------------------------------------------------------------
Hi,
For your reference, here is a summary of the transfer rates I've noticed
between two DECstation 5000/200, connected to a DECconcentrator 500. On all
the occasions the sender run out of cpu.
TCP/IP default 18 Mbps
UDP/IP default 20 Mbps
UDP/IP without checksum 34 Mbps
DECnet default 12 Mbps
DECnet SDU size 4357, segment buffer size 65535 14 Mbps
DECnet changed from NSP to OSI transport 17 Mbps
DECnet without congestion avoidance 19 Mbps
|
| 332.3 | | KONING::KONING | Brivu Latviju! | Thu Aug 29 1991 11:26 | 5 |
| I'm not sure how you run DECnet without congestion avoidance (other than by
patching the code) but in any case, note that doing so will hurt severely
as soon as you have others on the network along with you.
paul
|
| 332.4 | I used NCL on DECnet/Ultrix | STKHLM::TORGNY | | Thu Aug 29 1991 12:41 | 30 |
| Hi Paul,
as you see from the DECnet/Ultrix's ncl commands below, it's possible to change
the congestion avoidance attribute. Doing that, I noticed the performance
improvement that's listed in .2
Cheers,
Torgny...
ncl> show osi trans cong avoi
Node 0 OSI Transport
AT 1991-08-29-17:33:44.669+02:00I583.885
Characteristics
Congestion Avoidance = True
ncl> set osi tran cong avoi false
ncl> show osi trans cong avoi
Node 0 OSI Transport
AT 1991-08-29-17:33:55.438+02:00I583.886
Characteristics
Congestion Avoidance = False
|
| 332.5 | why selfish optimization works | DELNI::GOLDSTEIN | Networks designed while-u-wait | Thu Aug 29 1991 17:13 | 12 |
| re:.4
Yes, it may run faster if you turn congestion avoidance off...
YOUR node will run faster! But that's because it's "hogging" the
network. Others will slow down and you won't, so they'll be making way
for you. This is called "selfish optimization". It makes a good demo,
but a bad network. Total throughput amongst all nodes declines; the
amount you gain is less than others lose.
Of course, TCP/IP lacks congestion avoidance while DECnet V has it, so
I suppose it's fair to turn it off when the network is mostly TCP....
That's also one reason why IP fans haven't added it. Who'd go first?
|
| 332.6 | It was for a DEMO | STKHLM::TORGNY | | Fri Aug 30 1991 06:43 | 12 |
| > YOUR node will run faster! But that's because it's "hogging" the
> network. Others will slow down and you won't, so they'll be making way
> for you. This is called "selfish optimization". It makes a good demo,
> but a bad network. Total throughput amongst all nodes declines; the
> amount you gain is less than others lose.
Exactly, and I did this at a symposium on high speed communication and
I didn't want OSI to look bad compared with IP on the private FDDI ring
we used.
Torgny...
|
| 332.7 | particulars | BREAKR::HA | | Wed Sep 04 1991 18:47 | 13 |
| Here's what finally worked:
# dbx -k /vmunix
(dbx) assign tcpcksum=0
(dbx) assign tcp_sendspace=32768
(dbx) assign tcp_recvspace=32768
(dbx) quit
I now get 18-19 Mb/s using TCP. To boost UDP performance the variables
are called udpcksum, udp_sendspace, and udp_recvspace, respectively.
Michael
|