aboutsummaryrefslogtreecommitdiffstats
path: root/src/vnet/ip6-nd/ip6_nd_inline.h
diff options
context:
space:
mode:
authorFlorin Coras <fcoras@cisco.com>2022-11-18 18:29:23 -0800
committerDave Barach <openvpp@barachs.net>2022-11-28 19:31:59 +0000
commit797562c86066411c29116cc099864f12c98561d9 (patch)
tree828516d730a39080d37a9a62ba6e782ea28af695 /src/vnet/ip6-nd/ip6_nd_inline.h
parent7943c90b2e341e6e064538f31746e326cd7c4e42 (diff)
udp: preallocate ports sparse vec map
Not ideal. The sparse vector used to map ports to next nodes assumes only a few ports are ever used. When udp transport is enabled this does not hold and, to make matters worse, ports are consumed in a random order. This can lead to a lot of slow updates to internal data structures which in turn can slow udp connection allocations until all ports are eventually consumed. Consequently, reallocate sparse vector, preallocate all ports and have them point to UDP_NO_NODE_SET. We could consider switching the sparse vector to a preallocated vector but that would increase memory consumption for vpp deployments that do not rely on host stack. For reference, populating one of the v4 or v6 sparse vectors in reverse order takes about 9.8s on a skylake cpu. Type: improvement Signed-off-by: Florin Coras <fcoras@cisco.com> Change-Id: Id795e1805d0d3ba54f56a152a9506a7a2a06ecbc
Diffstat (limited to 'src/vnet/ip6-nd/ip6_nd_inline.h')
0 files changed, 0 insertions, 0 deletions