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(). > > 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? > | with SK_LOOKUP prog attached | 2,520,474 ± 0.7% | -12.1% | This also looks very different from udp4. > > Also visualized on bpf-sk-lookup-v1-rx-pps.png chart [2]. > > cpu-cycles measured with `perf record -F 999 --cpu 1-4 -g -- sleep 60`. > > | | cpu-cycles events | | > | tcp4 SYN flood | __inet_lookup_listener | Δ events | > |------------------------------+------------------------+----------| > | 5.6.3 vanilla (baseline) | 1.12% | - | > | no SK_LOOKUP prog attached | 1.31% | 0.19% | > | with SK_LOOKUP prog attached | 3.05% | 1.93% | > > | | cpu-cycles events | | > | tcp6 SYN flood | inet6_lookup_listener | Δ events | > |------------------------------+------------------------+----------| > | 5.6.3 vanilla (baseline) | 1.05% | - | > | no SK_LOOKUP prog attached | 1.68% | 0.63% | > | with SK_LOOKUP prog attached | 3.15% | 2.10% | > > | | cpu-cycles events | | > | udp4 0-len flood | __udp4_lib_lookup | Δ events | > |------------------------------+------------------------+----------| > | 5.6.3 vanilla (baseline) | 3.81% | - | > | no SK_LOOKUP prog attached | 5.22% | 1.41% | > | with SK_LOOKUP prog attached | 8.20% | 4.39% | > > | | cpu-cycles events | | > | udp6 0-len flood | __udp6_lib_lookup | Δ events | > |------------------------------+------------------------+----------| > | 5.6.3 vanilla (baseline) | 5.51% | - | > | no SK_LOOKUP prog attached | 6.51% | 1.00% | > | with SK_LOOKUP prog attached | 10.14% | 4.63% | > > Also visualized on bpf-sk-lookup-v1-cpu-cycles.png chart [3]. > [ ... ] > > [0] https://urldefense.proofpoint.com/v2/url?u=https-3A__blog.cloudflare.com_a-2Dtour-2Dinside-2Dcloudflares-2Dg9-2Dservers_&d=DwIDaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=VQnoQ7LvghIj0gVEaiQSUw&m=v4r30a5NaPFxNXVRakV9SeJkshbI4G4c5D83yZtGm-g&s=PhkIqKdmL12ZMD_6jY_rALjmO2ahv_KNF3F7TikyfTo&e= > [1] https://github.com/majek/inet-tool/blob/master/ebpf/inet-kern.c > [2] https://urldefense.proofpoint.com/v2/url?u=https-3A__drive.google.com_file_d_1HrrjWhQoVlqiqT73-5FeLtWMPhuGPKhGFX_&d=DwIDaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=VQnoQ7LvghIj0gVEaiQSUw&m=v4r30a5NaPFxNXVRakV9SeJkshbI4G4c5D83yZtGm-g&s=9tums5TZ16ttY69vEHkzyiEkblxT3iwvm0mFjZySJXo&e= > [3] https://urldefense.proofpoint.com/v2/url?u=https-3A__drive.google.com_file_d_1cYPPOlGg7M-2DbkzI4RW1SOm49goI4LYbb_&d=DwIDaQ&c=5VD0RTtNlTh3ycd41b3MUw&r=VQnoQ7LvghIj0gVEaiQSUw&m=v4r30a5NaPFxNXVRakV9SeJkshbI4G4c5D83yZtGm-g&s=VWolTQx3GVmSh2J7TQixTlGvRTb6S9qDNx4N8id5lf8&e= > [RFCv1] https://lore.kernel.org/bpf/20190618130050.8344-1-jakub@xxxxxxxxxxxxxx/ > [RFCv2] https://lore.kernel.org/bpf/20190828072250.29828-1-jakub@xxxxxxxxxxxxxx/
