1 files changed, 41 insertions, 0 deletions
diff --git a/results/classifier/118/performance/1312 b/results/classifier/118/performance/1312
new file mode 100644
index 000000000..c7a1810c1
--- /dev/null
+++ b/results/classifier/118/performance/1312
@@ -0,0 +1,41 @@
+performance: 0.984
+virtual: 0.922
+debug: 0.904
+device: 0.788
+kernel: 0.773
+files: 0.742
+network: 0.691
+ppc: 0.690
+permissions: 0.690
+vnc: 0.690
+PID: 0.677
+graphic: 0.672
+TCG: 0.627
+socket: 0.571
+hypervisor: 0.525
+semantic: 0.508
+risc-v: 0.497
+peripherals: 0.468
+arm: 0.464
+architecture: 0.455
+VMM: 0.417
+register: 0.414
+x86: 0.379
+boot: 0.310
+i386: 0.309
+mistranslation: 0.282
+user-level: 0.238
+assembly: 0.154
+KVM: 0.115
+
+TCP performance problems - GSO/TSO, MSS, 8139 related (Ignores lower MTU from PMTUD/MSS)
+Description of problem:
+MTU handling on guests using an RTL8139 virtualized NIC is broken; net/hw/8139.c works with a static MTU of 1500b for TCP offloading, leading to low throughput when clients connect from sub 1500MTU networks. PMTUD is ignored, and locking to a lower MTU in the OS mitigates the issue.
+Steps to reproduce:
+1. Create a guest with an RTL8139 nic
+2. Try to retrieve a file from a client behind a sub 1500 MTU link
+3. Observe low bandwidth due to retransmits
+Additional information:
+I just debugged this issue for an NGO which, for whatever reason, had an RTL8139 NIC in their guest. After i finally traced this to the RTL8139, i found this qemu-devel/netdev thread from six years ago, which apparently already debugged this issue and proposed a patch: https://lore.kernel.org/all/20161114162505.GD26664@stefanha-x1.localdomain/
+
+I did not test the patch proposed there, but note that `net/hw/8139.c` still looks as discussed in that qemu-devel/netdev thread. As i haven't found a bug report in the archives, i figured you might want to know.