beirne@richsun.cpg.trs.reuter.com (Michael G. Beirne) (06/04/90)
I have discovered two problems with the SPARC systems ethernet interfaces.
1. "spray localhost" loses about 10% of the packets sent. This doesn't
sound bad until you realize that localhost packets don't leave the system
and are just looped back in the kernel.
2. The SPARC systems can in certain circumstances put out bad ETHERNET
packets with both the destination and SOURCE addresses wrong.
Try this shell script:
for i in `ypcat hosts|awk '{print $1}'`;do ping $1 3;done
This will try to ping all the hosts in your NIS(yellow pages) hosts file.
Then use "etherfind -x -n -greater 60 -o -less 90 >/tmp/badpackets" The
bad packets have a type of 1F-00 which is unknown to Suns or anyone else
for that matter. This has been tried on a 4/110 running 4.0 and 4.0.1 and
also on a SPARCstation 1 running 4.0.3c and 4.1.
The first one would seem to mean that the SPARCstation is 10% artificially
slower on the ethernet due to dropping packets. Does this mean that if
this problem were fixed that NFS and telnet sessions would be 10% faster?
The second problem seems more serious, and may be the cause of the 10%
loss.
I reported the "spray localhost" problem last year and had hoped that it
would be fixed in 4.1 but no such luck. I put a call last week about the
second problem but SUN hasn't responded back yet.
Michael G. Beirne
beirne@richp1.UUCP or beirne@richsun.cpg.trs.reuter.com
or more reliably beirne@limerick.chi.il.usbeirne@richsun.reuter.com (Michael G. Beirne) (06/12/90)
In article <8507@brazos.Rice.edu> beirne@richsun.cpg.trs.reuter.com (Michael G. Beirne) writes: >I have discovered two problems with the SPARC systems ethernet interfaces. > >1. "spray localhost" loses about 10% of the packets sent. This doesn't > sound bad until you realize that localhost packets don't leave the system > and are just looped back in the kernel. This has been explained as not a problem as spray uses UDP/RPC and the SPARCstation is able to overrun the local ethernet buffers and since UDP is unreliable this is fine although misleading. >2. The SPARC systems can in certain circumstances put out bad ETHERNET > packets with both the destination and SOURCE addresses wrong. > for i in `ypcat hosts|awk '{print $1}'`;do ping $1 3;done This I have found is not a SPARCstation problem but a problem with several systems that use a very old and locally ported version of NRC's Fusion software. The people who did the port are no longer working for Rich/Reuters and the source was never put into SCCS so I guess I'll have to live with the problem. We do have a newer version of NRC's Fusion that does not have the same problem. I apologize to Sun for wrongly maligning their software. Michael G. Beirne beirne@richp1.UUCP or beirne@richsun.cpg.trs.reuter.com or more reliably beirne@limerick.chi.il.us