Dear vpp-devers, As I mentioned on the last community call, there is patch which significantly changes the way how VPP is allocating wired memory, including the dpdk hugepages.
Patch is available here and it is passing verify jobs: https://gerrit.fd.io/r/#/c/7701/ With this change, VPP is able dynamically pre-allocate hugepages, if they are not already available. This change is affecting DPDK buffer mempools, as now they are allocated by VPP, and not directly by DPDK. I use dpdk rte_mempool_xmem_create (...) to pass allocated wired memory region to dpdk. Result is smaller memory footprint, mainly thanks to the better control of memory allocation we have. DPDK is still allocating 64M/socket for it’s internal data structures and VPP allocates 40MB/socket for our default number of buffers (16K). In case people want more buffers, it is enough to increase num_mbufs parameter and VPP will increase size of mempool automatically, which is significant improvement as currently people need to play with socket-mem parameter. In total, footprint is reduced from 256M/socket to 104M/socket. At the moment, code only deals with 2M pages, support for 1G pages for extreme VPP consumers will be added in the separate patch. I will really appreciate if people can try this patch and report success/failure. It important that we test it in different configurations before it is merged. Thanks, Damjan _______________________________________________ vpp-dev mailing list vpp-dev@lists.fd.io https://lists.fd.io/mailman/listinfo/vpp-dev
