land v0.9.1 on HEAD, in preparation for a 1.0.x branch

[fs/lustre-release.git] / lnet / klnds / scimaclnd / README.scimacnal

diff --git a/lnet/klnds/scimaclnd/README.scimacnal b/lnet/klnds/scimaclnd/README.scimacnal
index d4c6a49..e1ee3b5 100644 (file)
--- a/lnet/klnds/scimaclnd/README.scimacnal
+++ b/lnet/klnds/scimaclnd/README.scimacnal
@@ -2,13 +2,41 @@
 scimacnal - A NAL for the Scali ScaMAC midlayer.
 
 The ScaMAC midlayer is a simplified API to the SCI high performance
-interconnect.
+interconnect (http://www.scali.com/, http://www.dolphinics.com/).
 
 In order to use this NAL you'll need to tune scimac to use larger buffers.
 See scimac.conf in this directory for an example.
 
-Overall performance and stability isn't great but this can be attributed
-to the scimac driver which apparently is in need of some development.
+You'll also need to edit portals/include/portals/lib-types.h and reduce
+the MTU to 64kB (the limit of scimac), this diff should help:
+
+----------------------8<----------------------------------
+--- portals/include/portals/lib-types.h 27 Jul 2003 02:05:47 -0000      1.1.2.5
++++ portals/include/portals/lib-types.h 14 Aug 2003 08:32:14 -0000
+@@ -137,8 +137,8 @@
+ } lib_counters_t;
+ 
+ /* temporary expedient: limit number of entries in discontiguous MDs */
+-# define PTL_MTU        (512<<10)
+-# define PTL_MD_MAX_IOV 128
++# define PTL_MTU        (64<<10)
++# define PTL_MD_MAX_IOV 16
+ 
+ struct lib_msg_t {
+         struct list_head  msg_list;
+----------------------8<----------------------------------
+
+The NAL itself seems quite stable, though scimac has recovery bugs when
+rebooting nodes at times (confirmed by the fact that the IP driver that
+also uses scimac loses connectivity when scimacnal does).  This is
+solved by unloading lustre and reloading the scimac driver on the
+affected nodes.
+
+Performance isn't great when it comes to latency, scimac seems to have
+problems with per packet latencies (confirmed with the IP driver which
+has similar behaviour). Bandwidth using large packets is pretty OK
+(probably due to the fact that it hides the latency issues).
 
 TODO:
 Routing isn't yet implemented.
+Need some way to inform portals about our MTU.