[I'm guessing that increasing the size of buf in RCS_copydeltas from
8192 to about 64k or 128k would probably fix this.  At least that is
the most obvious difference I saw between the CVS 1.10 code and the
RCS 5.7 code.  Of course, reading in the _whole file_ at once is a bad
idea.

Could people try this out and report back (to bug-cvs) whether it
seems to help?  At the risk of belaboring the obvious, I've included a
diff at the end of this message.

-kingdon]

The report:

  [10x to 100x slowdown in checkin, in CVS 1.10 as compared with RCS 5.7]

  The time is spent in the function RCS_copydeltas(), where the last
  while loop (it is the 'while(fread) { fwrite }' loop) is the problem.
  The actual throughput of this loop is only a couple of Kb's per second
  across my NFS [HPUX 10.20 NFS client, not sure about NFS server; also
  observed with Solaris 2.6 over NFS].

Index: rcs.c
===================================================================
RCS file: /home2/cvsroot/ccvs/src/rcs.c,v
retrieving revision 1.208
diff -c -r1.208 rcs.c
*** rcs.c	1998/09/25 01:25:18	1.208
--- rcs.c	1998/10/03 20:20:28
***************
*** 8210,8216 ****
      char *bufrest;
      int nls;
      size_t buflen;
!     char buf[8192];
      int got;
  
      /* Count the number of versions for which we have to do some
--- 8210,8216 ----
      char *bufrest;
      int nls;
      size_t buflen;
!     char buf[131072];
      int got;
  
      /* Count the number of versions for which we have to do some