[ntp:hackers] Patches to reinstate rank variable and add endkeyword.

Danny Mayer mayer at gis.net
Sun Aug 7 15:36:39 UTC 2005


Brad Knowles wrote:
> At 7:27 AM +0100 2005-08-07, David J Taylor wrote:
> 
>>  Many of today's present processors and perhaps most of tomorrow's 
>> will be
>>  either hyperthreading, dual-core or both.  Code which is not compatible
>>  this will be a pain.
> 
> 
>     Code based on eventlib will work just fine, without any changes, on 
> multi-core/multi-processor systems.  Code which is dependent on POSIX 
> threads (or other form of threading) will be relatively non-portable, 
> especially to older architectures.
> 
>     Since there should only ever be one process/thread that is updating 
> the system clock, using a non-threaded model for ntpd will work just 
> fine, so long as a little care is taken to handle multiprocessing with 
> select() or other appropriate calls.
> 

The advantage of using a multiprocessor and threads is that you can have 
  a different thread handling the I/O on a different processor as long 
as you have just one thread handling clock changes. It is preferable to 
have that thread stay on the one CPU. Some O/S's allow you to specify 
the CPU on which the thread is to run.

I seem to remember somewhere that a single CPU currently is better than 
a multiprocessor for ntpd. I think that PHK has done work on this.

Danny


More information about the hackers mailing list