Integrated collection of open PRs on master #424
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Strip away these options:
-DLUAJIT_USE_PERFTOOLS
This feature causes noise in /tmp (perf-nnnn.map files). There is a
better way to profile at trace granularity with the LuaJIT
profiler. Hopefully this will be upstream in LuaJIT but for now
available as a patch:
http://www.freelists.org/post/luajit/Profile-at-trace-granularity,1
-DLUAJIT_USE_GDBJIT
The documentation says that "enabling it always has a non-negligible
overhead -- do not use in release mode" and I don't believe that we
are using this functionality (debugging JIT traces in gdb).
-DLUAJIT_NUMMODE=3
This is less clear cut. Mike Pall has encouraged us to experiment with
settings for this variable for potential performance benefits. The
current value (3) was chosen during a micro-optimization spree many
moons ago. I am not confident whether this setting helps or is better
than other settings. So I am disabling it in this commit ("less
voodoo") and we will have to check for performance regressions before
merging onto master.
Simplify the code by operating on unaligned data. CPUs supporting AVX2 handle unaligned access very efficiently (source: Agner Fog). The previous version of this function used a non-SIMD checksum for any non-aligned data at the start of the packet. (This is still done for the non-aligned part at the end, since the AVX code still operates on minimum 32 byte chunks.)
This will make Snabb Switch always show 100% CPU in default settings. Previously Snabb Switch has by default attempted to bring in a batch of new packets once every 100us. This is too conservative on the straightline design because we now work on smaller batches of packets (max is now 256 and was 8192 before). So this change makes the default behavior to busy-loop without sleeping to maximize throughput. Going forwards we should consider a dynamic mechanism that will make a truly idle Snabb Switch sleep. Meanwhile the default is to busy-loop and a static configuration is possible with a setting like 'engine.Hz=10000'.
The standard link report printout is now a little prettier: links are alphabetically sorted (resolves snabbco#385), punctuation is improved, and values are printed with commas and justified. Before: link report 597223954 sent on Virtio_A.tx -> NIC_A.rx (loss rate: 0%)) 137411842 sent on NIC_A.tx -> Virtio_A.rx (loss rate: 0%)) 597223786 sent on NIC_B.tx -> Virtio_B.rx (loss rate: 0%)) 137411851 sent on Virtio_B.tx -> NIC_B.rx (loss rate: 0%)) After: link report: 17,747,602 sent on NIC_A.tx -> Virtio_A.rx (loss rate: 0%) 70,955,552 sent on NIC_B.tx -> Virtio_B.rx (loss rate: 0%) 70,955,729 sent on Virtio_A.tx -> NIC_A.rx (loss rate: 0%) 17,747,621 sent on Virtio_B.tx -> NIC_B.rx (loss rate: 0%)
Snabb Switch previously called mlockall() to force all memory in the process address space to be locked to its physical location. This made it possible to use all memory for DMA, including memory mapped from virtual machines. This had several side-effects: * Prevent Snabb Switch from being swapped out. (Nice?) * Prevent VMs that we serve from being swapped out. (Overkill?) * Requires root permissions at startup. (Overkill?) The new behavior is to individually mlock() the HugeTLB pages that we allocate for DMA. The rest of the memory - Snabb Switch and VMs - is left unlocked. This is feasible now that we only do DMA within this memory. This also allows us to defer the root-privileged system call until we need DMA memory, which may not happen for ceratin applications.
Wait until the first packet is needed to allocate the memory.
Certain Snabb programs don't require root permissions. The ones that do will need to check for them when system calls fail unexpectedly.
Use ljsyscall's geteuid() via a new library function:
lib.root_check()
Added proper usage for neutron-sync-master and neutron-sync-agent. Created separate usage commands for traffic: -h/--help prints command line usage, -H/--long-help also prints the configuration file format.
…le ethernet frame) and returns true if successufly checks header (if aplicable) and data.
It's faster to return a single value from C to Lua, even if that means recalculating the other one.
The optional arguments (only -B) were treated as positional arguments.
Conflicts: src/program/snabbnfv/traffic/traffic.lua
the generic loop takes the initial value in network order and swaps at the end, but the SIMD versions take it in host order.
Setting flags to C.VIO_NET_HDR_F_NEEDS_CSUM allowed even known-bad packets to pass. it's better to set it to 0 when we do have a failed checksum.
Tests that all checksum functions return the same value with the same parameters, including non-zero initial values.
lukego
added a commit
that referenced
this pull request
Mar 25, 2015
Integration branch merge containing PRs to master.
dpino
added a commit
to dpino/snabb
that referenced
this pull request
Sep 8, 2016
Load two virtual NICs if using two different VLAN tags
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Merge with testing of open pull requests to master:
mlock()less memory #417 Relax requirement to run as root andmlock()individual pages on demandsnabbnfv: Improve --help usage outputsnabbnfv traffic: Fix command-line argument parsing