[1922] in Kerberos_V5_Development
Re: Cygnus changes for your consideration
daemon@ATHENA.MIT.EDU (Ken Raeburn)
Thu Oct 31 17:16:55 1996
To: Ezra Peisach <epeisach@MIT.EDU>
Cc: "Theodore Y. Ts'o" <tytso@MIT.EDU>, Mark Eichin <eichin@cygnus.com>,
krbdev@MIT.EDU
From: Ken Raeburn <raeburn@cygnus.com>
Date: 31 Oct 1996 17:13:37 -0500
In-Reply-To: Ezra Peisach's message of Thu, 31 Oct 1996 13:13:44 EST
> The fast asn1 date encoders I think were attempts by Ken to take away
> from the bottlenecks in sending timestamps using essentially _doprnt internally.
It was actually a combination of things Mark and I came up with.
* Several changes got us from formatting a time string to picking
apart numbers and filling in bytes.
* A small 4- or 5-entry cache bypasses all the work for time values
recently converted. I think you'll tend to get more cache hits under
higher load. I'd have to look over the code to confirm it, but that
was the behavior I saw in my profiling tests.
* A few functions were made inline if gcc is used.
This reminds me -- there was some time field which was supposedly
optional, but our code was always encoding it, even when it was zero.
I even created a special non-expiring cache entry for it, though I
didn't initialize it. Did anyone ever figure out if it was safe to
omit it? (Not for 1.0, but for afterwards.) I recall some concern
about DCE compatibility. Bill, have you looked into that?
> The non-use of calloc I believe would be a win as we zero the memory when we
> need it anyway - so why do it twice.
In most of the cases I looked at, I think the memory was completely
overwritten, but not explicitly zeroed outside the calloc call. Still
amounts to redundant stores.
