i didnt work on the installation of AFS over ATM myself
(maybe the sysadmins at NRL can expand on the experiences
they had) - but i did look at the performance and the
code of AFS 3.3a to see if AFS was much faster over ATM.
the only problem with installation i do remember having
was it took us a while to realize that /etc/hostname on
the database server machines (all our machines were
multihomed) had to be the same thing listed in the
CellServDBs - i imagine that has been solved in AFS 3.4
here's the stuff i learnt from the internal changes:
(some of this might have changed with 3.4)
1. (as of AFS 3.3) if the ATM driver doesnt return the proper
MTU values in the ifnet structure - the transfer MTU defaults
to a lower (ethernet ?) value. this is what happened in our
case - so AFS did NOT take advantage of the ATM speeds.
2. after changing the code (at NRL) so that a different window
and mtu size was used depending on whether the connection was over
ATM or ethernet - i found the cache manager to be the bottleneck.
so even though the wire transfer rate went to about 19-25 Mbps
(i think it should go to much higher values than that and am
still hunting for other bottlenecks such as the OS socket buffer
size) - the application level (say, reading a file) saw 6 Mbps
at most.
3. Using a large chunksize (1M) also improved performance when
the connection was over ATM and using an MTU size of 8K (i
didnt use 9188)
Our initial performance numbers:
The experiment was simply reading a file. The time values
are measured at the application level (these are not the
times taken to transfer over the wire).
The ethernet measurements were based on unchanged transarc code.
The ATM columns were based on the changes i made here so far.
(Conceptually, the only real serious change thus far was larger
MTUs).
-----------------------------------------------------------
| File Size | Warm Cache | Cold Cache | Cold Cache |
| | | Ethernet | ATM |
| | | AFS 3.3 | AFS 3.3 |
| (MB) | (seconds) | (seconds) | (seconds) |
-----------------------------------------------------------
| 0,5 | 0.0 | 1.1 | 0.6 |
| 1.0 | 0.1 | 2.2 | 1.1 |
| 2.0 | 0.3 | 4.6 | 2.4 |
| 4.0 | 0.6 | 9.5 | 5.1 |
| 10.0 | 1.4 | 28.6 | 13.5 |
-----------------------------------------------------------
-sushi
