From: Hanna Linder <hannal@us...> - 2002-08-09 21:21:38
LSE Con Call Minutes from August 9, 2002
I. Mala Anand -
Mala found a problem with per cpu slab allocator implementation in Linux
kernel. The per cpu array of objects get emptied unnecessarily when
there is no room to put back the freed object.
Her skb patch moves the initialization so as not to initizalize
right before freeing the skb. Currently there is no guarantee
the object will be used again on the same CPU on an SMP system
so this patch avoids cacheline bouncing.
She found a 25% reduction in cycles with the patch.
II. Patrick Dempsey -
He has a 2-way and a 4-way at home and looked around for
patches to help his systems perform better. He discoverd
the lse-tech sourceforge site and decided it needed a
bitkeeper tree or two to help out.
So far the response has been positive and many people are
planning on using his trees in the future. He can be
reached at: bk@... The trees can be found on
III. Keith Owens -
A benefit of tihs is for multiple writes. currently only
used by sgi sn2. could be useful for numa. this is cool but it may be
a while before we will get to using it. havent posted yet to the
discontig mem mailing list. sgi also wrote the ia64 discontigmem
kernel text replication code.
IV. Craig Thomas -
OSDL has been trying to increase the number of users on
a database. noticed a response time degredation when number of users
increased. started with 2.4 now moving onto 2.5.30. having difficulties
getting 2.5.26 to work (have ide drives). did get a run on 2.5.30
had pretty bad performance numbers so far. pretty small system with
everything on one disk.
jenny - installed sapd rpm on system and doesnt work anymore,
basically fails to start the database. martin and dave also
said they found problems with no data at all coming through
apache so it may be a generic networking problem. the ones that
dont really use any networking work better. something may have
broke between 2.5.29 and 2.5.30 to affect networking stuff.
Cliff is going to try it on a different machine to see if they
can get the db up on it. We also have an 8-way we are trying
to get up on 2.4. It is a pentium III xeons. Dave said they
have been having some nmi problems and want to see if it
is reproducible running oprofile on this other xeon based
8-way. just let them know if it works or not. It might take
a couple weeks to get to that though.