> On 20 Aug 2018, at 23:44, Bly, Mike <m...@ciena.com> wrote: > > Hello, > > I am looking to understand the math and reasoning behind the 16K DPDK MBUFs > per socket allocation.
It is more educated guess than reasoning and math. > > From: vpp/src/plugins/dkdp/device/dpdk.h > > #define NB_MBUF (16<<10) > > The following online documentation has a comment stating it is 32K per > socket. Above == 16K, and does not seem to align with the 32K statement. > https://wiki.fd.io/view/VPP/Using_VPP_In_A_Multi-thread_Model > <https://wiki.fd.io/view/VPP/Using_VPP_In_A_Multi-thread_Model> Wiki i wrong, it should say 16K. I cannot update wiki, it tells me that I'm trying to do suspicious login :) > > Is there some basic math or VPP process understanding that can be shared? Total number of buffers needs to be bigger than all buffers waiting in the rx rings of all nics + number of buffers in flight during the one graph dispatch node iteration. I.e. in simple case with 2 DPDK NICSs, that will mean 1024 buffers on each ring + 2*256 buffers in-flight. Also at some places in the code we pre-allocate buffers for performance reasons so numbr might be slightly higher. > Is there an assumption as to the number of NICs or rx queues per NIC which > drives this #? yes, as explained above. > Are there assumptions being made relative to NIC port speed for this default > value? No, we have default of 1024 buffers on the RX ring for majority of DPDK NICs, For virtio number is 256. > Are there some online resources you can point me at to read and digest? Source code :) > > Regards, > Mike -- Damjan
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#10230): https://lists.fd.io/g/vpp-dev/message/10230 Mute This Topic: https://lists.fd.io/mt/24829323/21656 Group Owner: vpp-dev+ow...@lists.fd.io Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [arch...@mail-archive.com] -=-=-=-=-=-=-=-=-=-=-=-
