> > RIB nodes are allocated from a mempool created with no flags, so each > > object is cache-line aligned and padded for memory-channel spreading. > > On x86-64, an object takes 192 bytes for a 40-byte rte_rib_node or a > > 56-byte rte_rib6_node. Nodes are small and only accessed by the CPU, > > so the padding is wasted. The whole pool is reserved up front, and > > rte_fib and rte_fib6 size it at twice max_routes nodes. > > > > Create the node pools with RTE_MEMPOOL_F_NO_CACHE_ALIGN, > > RTE_MEMPOOL_F_NO_SPREAD and RTE_MEMPOOL_F_NO_IOVA_CONTIG. > > RTE_MEMPOOL_F_NO_SPREAD is implied by RTE_MEMPOOL_F_NO_CACHE_ALIGN, > > but is passed explicitly to document the intent. Nodes remain 8-byte > > aligned, which is all the node structures need. The extension area > > returned by rte_rib_get_ext() and rte_rib6_get_ext() was only 8-byte > > aligned before as well. > > > > With 2M nodes (1M routes), the pool shrinks from 390 MiB to 128 MiB > > for IPv4 and to 160 MiB for IPv6. On an AMD EPYC 7643 with 1 GiB > > hugepages, fib_perf_autotest (about 1M IPv4 routes) shows average add > > and delete times 13 and 29 percent lower. Lookup times, and the add > > and delete times in fib6_perf_autotest, which uses 1000 routes, are > > unchanged. > > > > Signed-off-by: Robert Sanford <[email protected]> > Acked-by: Vladimir Medvedkin <[email protected]>
Applied, thanks.
