On Mon, May 11, 2020 at 09:45 PM CEST, Martin KaFai Lau wrote: > On Mon, May 11, 2020 at 08:52:01PM +0200, Jakub Sitnicki wrote: > > [ ... ] > >> Performance considerations >> ========================== >> >> Patch set adds new code on receive hot path. This comes with a cost, >> especially in a scenario of a SYN flood or small UDP packet flood. >> >> Measuring the performance penalty turned out to be harder than expected >> because socket lookup is fast. For CPUs to spend >= 1% of time in socket >> lookup we had to modify our setup by unloading iptables and reducing the >> number of routes. >> >> The receiver machine is a Cloudflare Gen 9 server covered in detail at [0]. >> In short: >> >> - 24 core Intel custom off-roadmap 1.9Ghz 150W (Skylake) CPU >> - dual-port 25G Mellanox ConnectX-4 NIC >> - 256G DDR4 2666Mhz RAM >> >> Flood traffic pattern: >> >> - source: 1 IP, 10k ports >> - destination: 1 IP, 1 port >> - TCP - SYN packet >> - UDP - Len=0 packet >> >> Receiver setup: >> >> - ingress traffic spread over 4 RX queues, >> - RX/TX pause and autoneg disabled, >> - Intel Turbo Boost disabled, >> - TCP SYN cookies always on. >> >> For TCP test there is a receiver process with single listening socket >> open. Receiver is not accept()'ing connections. >> >> For UDP the receiver process has a single UDP socket with a filter >> installed, dropping the packets. >> >> With such setup in place, we record RX pps and cpu-cycles events under >> flood for 60 seconds in 3 configurations: >> >> 1. 5.6.3 kernel w/o this patch series (baseline), >> 2. 5.6.3 kernel with patches applied, but no SK_LOOKUP program attached, >> 3. 5.6.3 kernel with patches applied, and SK_LOOKUP program attached; >> BPF program [1] is doing a lookup LPM_TRIE map with 200 entries. > Is the link in [1] up-to-date? I don't see it calling bpf_sk_assign(). Yes, it is, or rather was. The reason why the inet-tool version you reviewed was not using bpf_sk_assign(), but the "old way" from RFCv2, is that the switch to map_lookup+sk_assign was done late in development, after changes to SOCKMAP landed in bpf-next. By that time performance tests were already in progress, and since they take a bit of time to set up, and the change affected just the scenario with program attached, I tested without this bit. Sorry, I should have explained that in the cover letter. The next round of benchmarks will be done against the now updated version of inet-tool that uses bpf_sk_assign: https://github.com/majek/inet-tool/commit/6a619c3743aaae6d4882cbbf11b616e1e468b436 > >> >> RX pps measured with `ifpps -d <dev> -t 1000 --csv --loop` for 60 seconds. >> >> | tcp4 SYN flood | rx pps (mean ± sstdev) | Δ rx pps | >> |------------------------------+------------------------+----------| >> | 5.6.3 vanilla (baseline) | 939,616 ± 0.5% | - | >> | no SK_LOOKUP prog attached | 929,275 ± 1.2% | -1.1% | >> | with SK_LOOKUP prog attached | 918,582 ± 0.4% | -2.2% | >> >> | tcp6 SYN flood | rx pps (mean ± sstdev) | Δ rx pps | >> |------------------------------+------------------------+----------| >> | 5.6.3 vanilla (baseline) | 875,838 ± 0.5% | - | >> | no SK_LOOKUP prog attached | 872,005 ± 0.3% | -0.4% | >> | with SK_LOOKUP prog attached | 856,250 ± 0.5% | -2.2% | >> >> | udp4 0-len flood | rx pps (mean ± sstdev) | Δ rx pps | >> |------------------------------+------------------------+----------| >> | 5.6.3 vanilla (baseline) | 2,738,662 ± 1.5% | - | >> | no SK_LOOKUP prog attached | 2,576,893 ± 1.0% | -5.9% | >> | with SK_LOOKUP prog attached | 2,530,698 ± 1.0% | -7.6% | >> >> | udp6 0-len flood | rx pps (mean ± sstdev) | Δ rx pps | >> |------------------------------+------------------------+----------| >> | 5.6.3 vanilla (baseline) | 2,867,885 ± 1.4% | - | >> | no SK_LOOKUP prog attached | 2,646,875 ± 1.0% | -7.7% | > What is causing this regression? > I need to go back to archived perf.data and see if perf-annotate or perf-diff provide any clues that will help me tell where CPU cycles are going. Will get back to you on that. Wild guess is that for udp6 we're loading and coping more data to populate v6 addresses in program context. See inet6_lookup_run_bpf (patch 7). This makes me realize the copy is unnecessary, I could just store the pointer to in6_addr{}. Will make this change in v3. As to why udp6 is taking a bigger hit than udp4 - comparing top 10 in `perf report --no-children` shows that in our test setup, socket lookup contributes less to CPU cycles on receive for udp4 than for udp6. * udp4 baseline (no children) # Overhead Samples Symbol # ........ ............ ...................................... # 8.11% 19429 [k] fib_table_lookup 4.31% 10333 [k] udp_queue_rcv_one_skb 3.75% 8991 [k] fib4_rule_action 3.66% 8763 [k] __netif_receive_skb_core 3.42% 8198 [k] fib_rules_lookup 3.05% 7314 [k] fib4_rule_match 2.71% 6507 [k] mlx5e_skb_from_cqe_linear 2.58% 6192 [k] inet_gro_receive 2.49% 5981 [k] __x86_indirect_thunk_rax 2.36% 5656 [k] udp4_lib_lookup2 * udp6 baseline (no children) # Overhead Samples Symbol # ........ ............ ...................................... # 4.63% 11100 [k] udpv6_queue_rcv_one_skb 3.88% 9308 [k] __netif_receive_skb_core 3.54% 8480 [k] udp6_lib_lookup2 2.69% 6442 [k] mlx5e_skb_from_cqe_linear 2.56% 6137 [k] ipv6_gro_receive 2.31% 5540 [k] dev_gro_receive 2.20% 5264 [k] do_csum 2.02% 4835 [k] ip6_pol_route 1.94% 4639 [k] __udp6_lib_lookup 1.89% 4540 [k] selinux_socket_sock_rcv_skb Notice that __udp4_lib_lookup didn't even make the cut. That could explain why adding instructions to __udp6_lib_lookup has more effect on RX PPS. Frankly, that is something that suprised us, but we didn't have time to investigate further, yet. >> | with SK_LOOKUP prog attached | 2,520,474 ± 0.7% | -12.1% | > This also looks very different from udp4. > Thanks for the questions, Jakub
