[j-nsp] About RE memory utilization

[Richard A Steenbergen]

On Fri, Apr 04, 2008 at 11:14:13AM -0400, Eric Van Tol wrote:
> Based only upon personal experience, we've had old M5s running at >95% 
> RE memory usage that hummed along just fine.  Even with FEB heap 
> utilization at around 90%, we didn't have problems.  Forwarding is done 
> in hardware, so you won't have a problem in that regard if memory usage 
> is too high.  However, FEB/SSB/cFEB memory can have an impact on route 
> updates, among other things.  Too high a utilization on there can 
> prevent route changes and other nastiness.

You can exceed your RE DRAM by quite a bit and still not break anything. 
There are literally hundreds of megs of memory pages for processes which 
aren't being used in realtime that will automatically be swapped out and 
almost never get swapped back in. The FreeBSD VFS is very good at handling 
this stuff, and swap acts as a great safety net against memory leaks since 
the leaked pages will be unreferenced and will be the first thing to get 
swapped out.

You can even get into situations where RPD is swapping parts of the 
routing table and not "break" break under normal use, just take a little 
longer to converge when something big happens. My first olive was a dinky 
p2 with 64MB of ram and even back then that didn't come close to fitting a 
full routing table, and the performance was surprisingly "not bad" 
considering half the routing table was being paged out. Hell I think it 
was faster back then running JUNOS 3.x and swapping than a RE-2.0 system 
running JUNOS 8.5 is today. :)

On the other hand, if you exceed your FEB DRAM you are 101% screwed.

-- 
Richard A Steenbergen <ras at e-gerbil.net>       http://www.e-gerbil.net/ras
GPG Key ID: 0xF8B12CBC (7535 7F59 8204 ED1F CC1C 53AF 4C41 5ECA F8B1 2CBC)