Dear Pavel, Thank you for your input. Please find my replies inline. Unfortunately, I will not be able to try your suggestions right away because I will not have access to my cluster for -- I guess -- a couple of weeks. I will update this thread of emails as soon as I have further news.
Regards, Harold 2017-07-24 18:23 GMT+02:00 Pavel Shirshov <pavel.shirs...@gmail.com>: > Hi Harold, > > Some thoughts: > > IPv4 packet_id. I'd better put packet_id in application header. > Otherwise a driver may interfere with packet. In case of Mellanox > there're two drivers: PMD + whole mellanox driver infrastructure. Try > to extend your application header with packet_id and test it. > *This is something I thought about several times, but never tried. I guess this will be the first thing I will try.* > > About registering missed packets. I'd put a switch between client and > server and start seeing into switch counters. It could give more > precise information about dropped packet and what component drops > them. In your case it's not clear your client or server dropped > packets. I don't believe both client and server statistics. > *I don't have direct access to the switch, unfortunately. Packets are counted as sent by the server in eth_stat, but not as received on the client.* > > Are you using RSS? What if you disable it? What if you enable it? Can > every second packet go to different NIC queue in high load? > *The load that I generate is not very high (it also happens with only two threads injecting traffic). RSS in general seems to be ok: if I only have single-frame requests, everything run smoothly.* *I have not tried disabling RSS yet because disabling it would cause frames from the same request to be received by different cores. This makes it harder to monitor exactly what is happening.* *However, it is something I shall try .* > > I don't think your static memory mbuf could mess anything. > *Yes, making my buffers dynamic didn't change anything* > > As I see you XEON has 8 cores/16 threads. Are you sure you're not > using threads on same cores? Because I see you have one cpu. All > cores/threads bound to the same NUMA domain. > > *AFAIK, lcores are automatically mapped to different cores at startup, so I don't "force" lcore to core affinity in any way. By spawning lcore 0..7 I expect them to be pinned to cores 0..7, which are in the same NUMA domain. * *Looking at htop, it seems this is fine.* > > On Mon, Jul 17, 2017 at 2:23 PM, Harold Demure > <harold.demur...@gmail.com> wrote: > > Dear Pavel, > > Thank you for your feedback; I really appreciate it. I reply to your > > questions inline. > > Regards, > > Harold > > > > 2017-07-17 22:38 GMT+02:00 Pavel Shirshov <pavel.shirs...@gmail.com>: > >> > >> Hi Harold, > >> > >> Sorry I don't have a direct answer on your request, but I have a bunch > >> of questions. > >> > >> 1. What is "packet_id" here? It's something inside of your udp payload? > > > > > > I have a packet_id in the plain ipv4_hdr structure and I have a > fragment_id > > in the header of each fragment I send > > So a typical packet is ETH|IP|UDP|APP_HEADER > > I defragment packets looking at the pckt_id in the ipv4 header and the > > fragment id in the app_header > > > >> 2. How do you know you have the packet loss? > > > > > > I know it because some fragmented packets never get reassembled fully. > If I > > print the packets seen by the server I see something like "PCKT_ID 10 > FRAG > > 250, PCKT_ID 10 FRAG 252". And FRAG 251 is never printed. > > > > Actually, something strange that happens sometimes is that a core > receives > > fragments of two packets and, say, receives frag 1 of packet X, frag 2 > of > > packet Y, frag 3 of packet X, frag 4 of packet Y. > > Or that, after "losing" a fragment for packet X, I only see printed > > fragments with EVEN frag_id for that packet X. At least for a while. > > > > This led me also to consider a bug in my implementation (I don't > experience > > this problem if I run with a SINGLE client thread). However, with smaller > > payloads, even fragmented, everything runs smoothly. > > If you have any suggestions for tests to run to spot a possible bug in my > > implementation, It'd be more than welcome! > > > > MORE ON THIS: the buffers in which I store the packets taken from RX are > > statically defined arrays, like struct rte_mbuf* temp_mbuf[SIZE]. SIZE > can > > be pretty high (say, 10K entries), and there are 3 of those arrays per > core. > > Can it be that, somehow, they mess up the memory layout (e.g., they > > intersect)? > > > >> > >> How can you be sure it's > >> packet loss if you don't see it on your counters? > > > > > > Just because i tag every packet, and some packets that should be there, > is > > not > > > >> > >> How can you be sure > >> that these packets were sent by clients? > > > > > > I print all packets that the clients send > > > >> How can you be sure your > >> clients actually sent the packets? > > > > > > The TX/mbuf error counters on the client eth_stats are 0 > > > > > >> > >> > >> Also I see you're using 2x8 cores server. So your OS uses some cores > >> for itself. Could it be a problem too? > >> > > > > I have no idea. I only use 8 cores out of the 16 I have, because I only > use > > the 8 that are in the same NUMA domain as the NIC. However, the PMD > should > > take the NIC out of the control of the kernel, so the OS should not be > able > > to see it or mess with it. > > > >> > >> Thanks > >> > >> On Mon, Jul 17, 2017 at 6:18 AM, Harold Demure > >> <harold.demur...@gmail.com> wrote: > >> > Hello, > >> > I am having a problem with packets loss and I hope you can help me > >> > out. > >> > Below you find a description of the application and of the problem. > >> > It is a little long, but I really hope somebody out there can help me, > >> > because this is driving me crazy. > >> > > >> > *Application* > >> > > >> > I have a client-server application; single server, multiple clients. > >> > The machines have 8 active cores which poll 8 distinct RX queues to > >> > receive > >> > packets and use 8 distinct TX queues to burst out packets (i.e., > >> > run-to-completion model). > >> > > >> > *Workload* > >> > > >> > The workload is composed of mostly single-frame packets, but > >> > occasionally > >> > clients send to the server multi-frame packets, and occasionally the > >> > server > >> > sends back to the client multi-frame replies. > >> > Packets are fragmented at the UDP level (i.e., no IP fragmentation, > >> > every > >> > packet of the same requests has a frag_id == 0, even though they share > >> > the > >> > same packet_id). > >> > > >> > *Problem* > >> > > >> > I experience huge packet loss on the server when the occasional > >> > multi-frame > >> > requests of the clients correspond to a big payload ( > 300 Kb). > >> > The eth stats that I gather on the server say that there is no error, > >> > nor > >> > any packet loss (q_errors, imissed, ierrors, oerrors, rx_nombuf are > all > >> > equal to 0). Yet, the application is not seeing some packets of big > >> > requests that the clients send. > >> > > >> > I record some interesting facts > >> > 1) The clients do not experience such packet loss, although they also > >> > receive packets with an aggregate payload of the same size of the > >> > packets > >> > received by the server. The only differences w.r.t. the server is > that a > >> > client machine of course has a lower RX load (it only gets the replies > >> > to > >> > its own requests) and a client thread only receives packets from a > >> > single > >> > machine (the server). > >> > 2) This behavior does not arise as long as the biggest payload > exchanged > >> > between clients and servers is < 200 Kb. This leads me to conclude > that > >> > fragmentation is not te issue (also, if I implement a stubborn > >> > retransmission, eventually all packets are received even with bigger > >> > payloads). Also, I reserve plenty of memory for my mempool, so I don't > >> > think the server runs out of mbufs (and if that was the case I guess I > >> > would see this in the dropped packets count, right?). > >> > 3) If I switch to the pipeline model (on the server only) this problem > >> > basically disappears. By pipeline model I mean something like the > >> > load-balancing app, where a single core on the server receives client > >> > packets on a single RX queue (worker cores reply back to the client > >> > using > >> > their own TX queue). This leads me to think that the problem is on the > >> > server, and not on the clients. > >> > 4) It doesn't seem to be a "load" problem. If I run the same tests > >> > multiple > >> > times, in some "lucky" runs I get that the run-to-completion model > >> > outperforms the pipeline one. Also, the run-to-completion model with > >> > single-frame packets can handle a number of single-frame packets per > >> > second > >> > that is much higher than the number of frames per second that are > >> > generated > >> > with the workload with some big packets. > >> > > >> > > >> > *Question* > >> > > >> > Do you have any idea why I am witnessing this behavior? I know that > >> > having > >> > fewer queues can help performance by relieving contention on the NIC, > >> > but > >> > is it possible that the contention is actually causing packets to get > >> > dropped? > >> > > >> > *Platform* > >> > > >> > DPDK: v 2.2-0 (I know this is an old version, but I am dealing with > >> > legacy code I cannot change) > >> > > >> > MLNX_OFED_LINUX-3.1-1.0.3-ubuntu14.04-x86_64 > >> > > >> > My NIC : Mellanox Technologies MT27520 Family [ConnectX-3 Pro] > >> > > >> > My machine runs a 4.4.0-72-generic on Ubuntu 16.04.02 > >> > > >> > CPU is Intel(R) Xeon(R) CPU E5-2630 v3 @ 2.40GHz 2x8 cores > >> > > >> > > >> > Thank you a lot, especially if you went through the whole email :) > >> > Regards, > >> > Harold > > > > >
