Datacenter applications demand microsecond-scale tail
latencies and high request rates from operating
systems. Achieving these goals in a CPU-efficient way
is an open problem. Because of the high overheads of
today’s kernel, the best available solution to achieve
microsecond-scale latencies is kernel-bypass networking,
which dedicates CPU cores to applications for spin-polling
the network card. But this approach wastes CPU: even at
modest average loads, one must dedicate enough cores for
the peak expected load.
In this talk Amy Ousterhout will describe the
approach taken by Shenango which achieves much
better CPU efficiency while maintaining the high
request rates.
More info:
https://netdevconf.org/0x13/session.html?talk-shenango
A reminder to all:
Early bird registration is still open until Feb 20th.
https://netdevconf.org/0x13/registration.html
cheers,
jamal
NutsBolts Talk.
As of kernel 4.18 the virtio guest driver uses the
SRIOV VF channel when it can or otherwise falls back
to the classical the para-virtual channel. With this
setup live-migration of VMs running under SRIOV
is possible.
Or Gerlitz and Parav Pandit handle the v-switch
host side of the equation for live-migration.
They propose a design based on the switchdev mode for
NIC host drivers.
More info:
https://netdevconf.org/0x13/session.html?talk-v-switch-virtio-sriov-vf-data…
A reminder to all:
Early bird registration is still open until Feb 20th.
https://netdevconf.org/0x13/registration.html
cheers,
jamal
TCP slow start faces a dilemma: Either you take too
long to ramp up or you ramp up too fast and cause
significant queuing delays. Joakim Misund et al
defined Paced Chirping which gets a flow to achieve
fast acceleration with virtually no queuing delay
The implementation extends the kernel's pacing framework
to allow a congestion control module to create "chirps"
with desired characteristics. Joakim will describe
the implementation and how one would use the extensions.
More info:
https://netdevconf.org/0x13/news.html?talk-chirp
A reminder to all:
Early bird registration is still open until Feb 20th.
https://netdevconf.org/0x13/registration.html
cheers,
jamal
Viet-Hoang Tran and Olivier Bonaventure leverage
Lawrence Brakmo's eBPF TCP-BPF framework to
allow TCP options extensions. Users can attach
eBPF code to inject and consume TCP options.
The authors will describe the code architecture
(currently on top of Kernel 4.17-rc5), illustrate
their use cases and finally provide performance
numbers.
More info:
https://netdevconf.org/0x13/session.html?talk-tcp-ebpf
A reminder to all:
Early bird registration is still open until Feb 20th.
https://netdevconf.org/0x13/registration.html
cheers,
jamal
Low-Power Wide Area Network (LPWAN) wireless
technologies sacrifice data throughput for long-range
communication with to preserve battery lifetime.
Andreas Färber created a design to bring LoRa PHY sockets
into the Linux kernel, with LoRaWAN MAC layered on top.
As he interacted with other LPWA or WPA network technologies
stake holders it was natural for more discussions
to surface.
This talk will give an update on where the discussions
have gone for LoRa, FSK, etc., with focus on netlink
layer, protocol families and socket addresses. Andreas
hopes to elicit more discussions and reach a consensus.
More info:
https://www.netdevconf.org/0x13/news.html?talk-ulpwa
A reminder to all:
Early bird registration is still open until Feb 20th.
https://netdevconf.org/0x13/registration.html
cheers,
jamal
The Identifier-Locator Network Protocol (ILNP) is
defined as an Experimental Internet Protocol by the
Internet Research Task Force (IRTF) in RFCs 6740-6748.
At the heart of the ILNP architecture is the desire
to address the deprecation of IP addresses. ILNP
replaces IP address semantics with use of node
Identifiers and network Locators using IPv6. No
application changes needed.
Current kernel implementation is on top of 4.19.
Ryo and Saleem plan to upgrade to a newer kernel
and push upstream. In this talk they will describe
the ILNP architecture, the kernel implementation of
ILNP and include results of testbed experiments for
IP-level mobility.
More info:
https://www.netdevconf.org/0x13/session.html?talk-ilnp
cheers,
jamal
Network applications can benefit from reduced CPU
cycles by amortizing the system call overhead of
network I/O operations.
In this talk Rahul et al review two existing
interfaces for network I/O batching namely
recvmmsg()/sendmmsg() and SO_RCVLOWAT and then
propose extensions to these mechanisms.
We show an 8x syscall reduction with our traffic
patterns (gaming scenarios) with the usage of such
extensions.
More info:
https://netdevconf.org/0x13/session.html?talk-syscall-batch
A reminder to all:
Early bird registration is still open until Feb 20th.
https://netdevconf.org/0x13/registration.html
cheers,
jamal
Hajime Tazaki is a man on a mission. We have seen
him before talking about the Linux Kernel Library (LKL)
whose intention was to take the whole of the linux
kernel network stack and make it available in
user space. He has since been staring at other possible
candidates for user space network stacks.
Hajime tries to come up with some metrics
on how to measure "maturity" of network stack
implementations: Using network protocol conformance tests
of various IETF standards (RFCs) across multiple userspace
network stack implementations.
In this talk, he is going to share the preliminary results,
findings of buggy implementations, and possible testing
framework that is going to be used going forward
https://netdevconf.org/0x13/session.html?talk-reimp-stack
A reminder to all:
Early bird registration is still open until Feb 20th.
https://netdevconf.org/0x13/registration.html
cheers,
jamal
Over the last few years, NIC speeds have gone up
from 1G to 10G to 25G to 40G and now 100G and 200G
and even 400G NICS are going to be showing up in
the commodity market soon.
Every time these rates go up, new challenges come
to the surface in the kernel. It is imperative to
understand those issues and resolve them.
In this talk Tariq Toukan will review the main bottlenecks
that were/are being addressed to support NIC speeds of
100Gbps, 200Gbps, and beyond.
He will summarize several ideas and solutions ranging
from traditional tricks of the trade like page recycling
to innovative HW capabilities that helped to recently
achieve milestone packet rate of 100 Million PPS in
XDP_TX and 135Mpps in XDP_DROP.
More info:
https://netdevconf.org/0x13/session.html?talk-bottlenecks
A reminder to all:
Early bird registration is still open until Feb 20th.
https://netdevconf.org/0x13/registration.html
cheers,
jamal
The Linux kernel has successfully made its way across
datacenter top-of-rack and spine switches.
David Lamparter says the data centre is a friendly
sane environment. You aint seen nothing yet.
Campus distribution and access switches need
to deal with a much more hostile environment.
In this talk he intends to collate the currently relevant
features for deploying campus and access environments.
He will give some operational context, and suggest
implementation strategies.
https://netdevconf.org/0x13/session.html?talk-linux-switch-req-feat
A note to all:
Early bird registration is still open until Feb 20th.
https://netdevconf.org/0x13/registration.html
cheers,
jamal