On Fri, Sep 02, 2016 at 06:59:21PM +0200, Alexander Bluhm wrote: > Hi, > > To move our network performance to modern high bandwith and high > latency characteristics, we have to increase the socket buffer size > limit. That also implies more mbuf clusters to avoid running out > of them. > > This diff includes several aspects that somehow belong together. > > - Increase the number of mbufs on most architectures. I have to > guess how much memory a typical machine has. If the value is too > high, we may run out of kernel memory. > arch/alpha/include/param.h:#define NMBCLUSTERS (16 * 1026)
1024 ? ---------------------------------------------------------^ > =================================================================== > RCS file: /data/mirror/openbsd/cvs/src/sys/arch/alpha/include/param.h,v > retrieving revision 1.39 > diff -u -p -r1.39 param.h > --- arch/alpha/include/param.h 26 Mar 2013 05:04:08 -0000 1.39 > +++ arch/alpha/include/param.h 2 Sep 2016 15:35:35 -0000 > @@ -63,7 +63,7 @@ > #define USPACE (UPAGES * PAGE_SIZE) /* total size of u-area > */ > #define USPACE_ALIGN 0 /* u-area alignment > 0-none */ > > -#define NMBCLUSTERS 4096 /* map size, max > cluster allocation */ > +#define NMBCLUSTERS (16 * 1026) /* max cluster > allocation */ ditto