Network stack personality in Android phone - netdev 2.2
2 likes1,149 views
An alternate approach to updating the Android network stack is to implement the stack in userspace using the Linux Kernel Library (LKL). This allows new network protocols and stacks to be added without replacing the host kernel. Measurements show the userspace LKL stack can achieve comparable performance to the native kernel over WiFi, and higher throughput over multipath TCP despite processing packets twice. However, results are often unstable and there are some limitations like asynchronous signal safety that require further work. Overall, using LKL provides a way to more easily distribute new network code without waiting for kernel upgrades.
![11
HOWTO
% LD_PRELOAD=liblkl-hijack.so netperf XXX # console app
% setprop wrap.app LD_PRELOAD=liblkl-hijack.so # Java app
{
"gateway": "10.206.211.1",
"interfaces": [
{
"ifgateway": "202.214.86.129",
"ip": "202.214.86.168",
"mac": "02:87:f8:27:22:02",
"masklen": "26",
"param": "/dev/tap23",
"type": "macvtap"
}
],
"debug": "0",
"singlecpu": "1",
"delay_main": "500000",
"sysctl": "net.ipv4.tcp_wmem=4096 87380 2147483647;net.mptcp.mptcp_debu
}](https://image.slidesharecdn.com/netdev22-tazaki-171108042212/85/Network-stack-personality-in-Android-phone-netdev-2-2-11-320.jpg)
![26
Alternate network stacks
lwip (2002~)
mTCP [NSDI '14]
SandStorm [SIGCOMM '14]
rumpkernel [ATC '09]
SolarFlare (2007~?)
libuinet (2013~)
SeaStar (2014~)
None of them are feature-rich, or one-shot porting](https://image.slidesharecdn.com/netdev22-tazaki-171108042212/85/Network-stack-personality-in-Android-phone-netdev-2-2-26-320.jpg)
Network stack personality in Android phone - netdev 2.2
- 1. 1 Network stack personality in Android phone Cristina Opriceana, Hajime Tazaki (IIJ Research Lab.) Linux netdev 2.2, Seoul, Korea 08 Nov. 2017
- 2. 2 Librarified Linux taLks (LLL) Userspace network stack (NUSE) in general (netdev0.1) kernel CI with libos and ns-3 (netdev1.1) Network performance improvement of LKL (netdev1.2, by Jerry Chu) How bad/good with LKL and hrtimer (BBR) (netdev2.1) Updating Android network stack (netdev2.2)
- 3. 3 Android a platform of billions devices billions installed Linux kernel Questions When our upstreamed code available ? What if I come up with a great protocol ? https://developer.android.com/about/dashboards/index.html
- 4. 4 Android (cont'd) When our upstreamed code available ? wait until base kernel is upgraded backport specific function What if I come up with a great protocol ? craft your own kernel and put into your image Long delivery to all billions devices
- 5. Approaches to alleviate the issue Virtualization (KVM on Android) Overhead isn't negligible to embedded devices Project Treble (since Android O) More modular platform implementation Fushia Rewrite OS from scratch QUIC (transport over UDP) Rewrite transport protocols on UDP
- 6. 5 https://source.android.com/devices/architecture/treble 6 An alternate approach network stack personality use own network stack implemented in userspace no need to replace host kernels but (try to) preserve the application compatibility NUSE (network stack in userspace) No delay of network stack update Application can choose a network stack if needed
- 7. 7 Userspace implementations Toys, Misguided People Selfish Motivation Trying to present that a Toy is practically useful
- 8. 8 Linux Kernel Library intro (again) Out-of-tree architecture (h/w-independent) Run Linux code on various ways with a reusable library h/w dependent layer on Linux/Windows /FreeBSD uspace, unikernel, on UEFI, network simulator (ns-3) Android
- 9. 9 LKL: current status Sent RFC (Nov. 2015) no update on LKML since then have evolved a lot fast syscall path offload (csum, TSO/LRO) CONFIG_SMP (WIP) json config qemu baremetal (unikernel) on UEFI https://github.com/lkl/linux
- 10. 10 Extensions to LKL Android (arm/arm64) support (lkl/linux#372) raw socket extension (only handle ETH_P_IP) (not upstreamed yet) hijack library enhance (not upstreamed yet)
- 11. 11 HOWTO % LD_PRELOAD=liblkl-hijack.so netperf XXX # console app % setprop wrap.app LD_PRELOAD=liblkl-hijack.so # Java app { "gateway": "10.206.211.1", "interfaces": [ { "ifgateway": "202.214.86.129", "ip": "202.214.86.168", "mac": "02:87:f8:27:22:02", "masklen": "26", "param": "/dev/tap23", "type": "macvtap" } ], "debug": "0", "singlecpu": "1", "delay_main": "500000", "sysctl": "net.ipv4.tcp_wmem=4096 87380 2147483647;net.mptcp.mptcp_debu }
- 12. 12 hijack library For smooth replacement (i.e., hijack) for Android UI app syscalls (java-based) bionic is more familiar than glibc only socket-related calls are redirected handling a mixture of host and lkl descriptors
- 13. 13 New feature introduction Example Multipath TCP ( ) out-of-tree for long time http://multipath-tcp.org/
- 14. 14 Multipath TCP An extension to TCP subsystem application compatibility (unlike SCTP) Use multiple paths better throughput (aggregation) smooth recovery from failure (handover) http://blog.multipath-tcp.org/blog/html/2015/12/25/commercial_usage_of_multipath_tcp.html
- 15. 15 Demo verify site (cat /proc/net/mptcp base detection) http://amiusingmptcp.de/
- 16. 16 No penalty with userspace network stack ? Condition To use Linux mptcp w/o replacing kernel Questions Is NUSE working fine (Will users wanna use it) ? How different from native Linux kernel ? With tolerable amount of overhead ?
- 17. 17 netperf measurement Client Nexus5 anrdoid 6.01 (rooted) LTE, wifi LKL arm/android patched or native kernel Server Ubuntu 16.04 (amd64) on KVM virtio/Etherlink (uplink: 100 Mbps) mptcp-4.4.70 (v0.92) Software netperf 2.7.x 10 seconds TCP_STREAM, TCP_MAERTS 5 trials, over 64-64K byte packet
- 18. 18 Single path (Wi-Fi only) Tx (TCP_STREAM) Rx (TCP_MAERTS) Condition phone: LKL v.s. (stock) kernel Comparable goodput CPU utilization: LKL < native
- 19. 19 Multipath TCP Tx (TCP_STREAM) Rx (TCP_MAERTS) Condition phone: LKL v.s. mptcp kernel Goodput (Tx) LKL > native even it's using multipath CPU: unstable (LKL) LKL > native
- 20. 20 Multipath TCP (Korea/KT) Tx (TCP_STREAM) Rx (TCP_MAERTS) Condition phone: LKL v.s. (stock) kernel native uses single-path/ LKL uses multi-path at Ibis hotel Goodput: No much gain with LKL even it's using multipath CPU: unstable (LKL) LKL > native
- 21. 21 Observations IP conflicts may heavier processed twice (host/lkl) per packet Results are often unstable difficult measurement under wireless media
- 22. 22 Limitations Implementations DHCP only boot time (handover will fail) IPv4 only on cellular interface (rmnet0) Fundamental limitations of hijack library asynchronous signal unsafe MT unsafe Required tweaks grant NET_RAW permission (packet socket) need filter out RST packet from host iptables -A OUTPUT -p tcp --tcp-flags RST RST -j DROP
- 23. 23 Further investigations other platform iOS11 now shipped userspace implementation profiling
- 24. 24 Summary Use out-of-tree kernel as a library on Android make your code easier to distribute with privileged installation/operation Comparable goodput over WiFi/LTE Unstable CPU utilization with LKL You can prepare your library file for your own purpose
- 25. 25 Backups
- 26. 26 Alternate network stacks lwip (2002~) mTCP [NSDI '14] SandStorm [SIGCOMM '14] rumpkernel [ATC '09] SolarFlare (2007~?) libuinet (2013~) SeaStar (2014~) None of them are feature-rich, or one-shot porting