There is no such thing as a free lunch. If you force a fresh allocation measured in gigabytes, the memory allocator will mmap(...) a bunch of (4k) pages which will incur (expensive) pagefaults as they’re populated. Vec_validate(...) copies data when necessary. When structures grow to gigabytes, that takes a measurable amount of time and will almost certainly result in packet drops.
Preallocation is the only strategy for avoiding packet drops when data structures expand beyond a certain size. Either that, or simply accept vector-size excursions until vpp reaches steady state: at that point, data structures shouldn’t have to expand, and you shouldn’t drop any more packets. Note that vec_reset_length(v) is highly preferable to constant cycling of fresh large vectors from vec_validate(...) / vec_add1(...) to vec_free(...). HTH... D. From: vpp-dev@lists.fd.io <vpp-dev@lists.fd.io> On Behalf Of david.leitch....@gmail.com Sent: Wednesday, October 24, 2018 11:06 AM To: vpp-dev@lists.fd.io Subject: Re: [vpp-dev] Memory Performance issue #vpp [Edited Message Follows] Do you mean it is impossible to have packet processing and memory operation at the same time, for example doing vec_validate or vec_free when NAT plugin is working and create new session. I have drop rate when vec_free or vec_vlidate for memory size greater than 3GB. What are your suggestions for such problems?
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#10961): https://lists.fd.io/g/vpp-dev/message/10961 Mute This Topic: https://lists.fd.io/mt/27615950/21656 Mute #vpp: https://lists.fd.io/mk?hashtag=vpp&subid=1480452 Group Owner: vpp-dev+ow...@lists.fd.io Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [arch...@mail-archive.com] -=-=-=-=-=-=-=-=-=-=-=-
