Hi, Thanks for the help, we will look into the suggested and get back if we have any concerns.
Thanks Pragash Vijayragavan pxv3...@rit.edu 585 764 4662 On Sat, Feb 18, 2017 at 4:12 PM, Dave Barach (dbarach) <dbar...@cisco.com> wrote: > A bit more inline, see drb>>> > > Thanks… Dave > > -----Original Message----- > From: vpp-dev-boun...@lists.fd.io [mailto:vpp-dev-boun...@lists.fd.io] On > Behalf Of Neale Ranns (nranns) > Sent: Saturday, February 18, 2017 3:52 PM > To: pxv3...@rit.edu; vpp-dev@lists.fd.io > Cc: Shailesh Vajpayee <srv6...@rit.edu>; Arjun Dhuliya <amd5...@rit.edu>; > Minseok Kwon <mxk...@rit.edu> > Subject: Re: [vpp-dev] vpp query > > > Hi Pragash, > > Some answers inline @[NR] > > > From: <vpp-dev-boun...@lists.fd.io> on behalf of Pragash Vijayaragavan < > pxv3...@rit.edu> > Reply-To: "pxv3...@rit.edu" <pxv3...@rit.edu> > Date: Saturday, 18 February 2017 at 03:56 > To: "vpp-dev@lists.fd.io" <vpp-dev@lists.fd.io> > Cc: Shailesh Vajpayee <srv6...@rit.edu>, Arjun Dhuliya <amd5...@rit.edu>, > Minseok Kwon <mxk...@rit.edu> > Subject: [vpp-dev] vpp query > > Hi, > > > We are working on a bloom Filter algorithm for ipv6 lookup in vpp and we > have a few queries. > It would be great if someone could help us with the following queries. > > 1. from the code, ip6_forward.c file processes 2 packets at a time from > the frame. > how are threads used in this. Is there a thread for each frame to be > processed or a thread per packet. > > [NR] there is one thread per frame. If multiple threads are used, then > each thread will invoke ip6_lookup_inline will a different frame. > Processing two packets at once in the same thread is to improve d-cache > performance. > > drb>>> By default, the dual-loop pattern prefetches 2 packets ahead. It > sometimes helps a bit to change the prefetch stride, e.g. to 3 or 4 ahead. > Vpp runs with a measured instructions-per-clock > 2. Until the compiler > runs out of registers, processing multiple packets at once helps the > hardware exploit fine-grained parallelism. Simple story: packet i uses a > set of registers to do <whatever>. Packet i+1 uses a disjoint set of > registers to do <whatever>. > > Selected routines benefit from quad-looping, processing 4 packets at a > time. Take a look at dpdk_device_input(...) in .../src/vnet/devices/dpdk/ > node.c. > > 2. A problem which we came across was synchronization for the lookup and > filling of the filter, Dave suggested > we use 2 filters and swap between them to address this issue > Is there any limitation on the amount of memory to be used for the filter. > > [NR] the only limit is the heap size. This is (1<<30) bytes by default. > You can change it; > https://wiki.fd.io/view/VPP/Command-line_Arguments#. > 22heapsize.22_parameter > > drb>>> Note that if you need a heap > 4gb, special arrangements are > required. Ping me if you think you need to go there. > > 3. Also we are required to handle 2 threads for our filter, one is to fill > up the filter using the fib entries > and the other is for lookup > - the ip6_forward.c -> ip6_lookup_inline function does the lookup part. > From our understanding this function > is processing the packets and checking if the destination ip is in the > fib. But we are required to fill our filter > with all fib entries dynamically as well, we understand this is also > similar to the mtrie code, but we are not able to > get how it works and threading is done here. > > [NR] > there’s a brief into the thread models here: > https://wiki.fd.io/view/VPP/Using_VPP_In_A_Multi-thread_Model > > in multi-thread mode worker threads invoke ip6_lookup_inline, the main > thread does the filing. In single thread mode the main thread does both. > > The IPv6 FIB is composed of 2 tables (where a table in this context means > a DB keyed by prefix); forwarding and non-forwarding. > The non-forwarding table contains ALL the prefixes that VPP is aware of – > a lookup in this table will result in a fib_entry_t. > The forwarding table contains the sub-set of prefixes that can be used for > forwarding – a lookup in this table will result in a load_balance_t. > I suspect you want to use bloom filters for the forwarding table. > Additions and removals from this table are via ip6_fib_table_fwding_dpo_update > and ip6_fib_table_fwding_dpo_remove respectively. > > HTH, > neale > > -- > Thanks, > > Pragash Vijayaragavan (pxv3...@rit.edu) > ph : 585 764 4662 > > > > _______________________________________________ > vpp-dev mailing list > vpp-dev@lists.fd.io > https://lists.fd.io/mailman/listinfo/vpp-dev > -- Thanks, Pragash Vijayaragavan (pxv3...@rit.edu) ph : 585 764 4662
_______________________________________________ vpp-dev mailing list vpp-dev@lists.fd.io https://lists.fd.io/mailman/listinfo/vpp-dev
