I’ve just shared the 2022-04 update in the Shadow forum: What's happening in Shadow 2022-04 · Discussion #2007 · shadow/shadow · GitHub
Also it looks like I forgot to share the previous 2021-12 update here: What's happening in Shadow 2021-12 · Discussion #1824 · shadow/shadow · GitHub
Both mirrored below:
We are continuing to work on Shadow 2.1. The biggest user-facing goal for this release is to support running golang programs in Shadow, especially tor simulations using the snowflake pluggable transport.
We’ve also begun planning the Shadow 2.2 release, which will largely be a push to refactor and migrate more of the core Shadow code to Rust.
- Optionally move time forward in non-blocking syscalls. Historically Shadow doesn’t move time forward except when explicitly waiting for an event, such as for a deadline to pass or for data to arrive over the network. Conceptually, this emulates a system with an infinite number of infinitely fast CPUs. Normally this is sufficient for modeling networks where CPUs aren’t expected to be a bottleneck. Unfortunately as we’ve explored running more software under Shadow, we’ve found a growing number of examples of code with “busy loops”, which deadlock in this model. This new feature optionally models every syscall taking some small amount of time (e.g. a microsecond), which allows the simulation to escape most such loops. We are still testing and improving this feature, and expect some version of it to be enabled by default in the next release.
- Fixed a corner case in getaddrinfo.
- Fixed several bugs in handling file descriptors:
Reliably intercept time via vdso. Previously we relied on intercepting calls to VDSO functions (such as
gettimeofday) to be intercepted at the libc level via
LD_PRELOAD. However, this doesn’t work when the VDSO is used more directly, such as in golang, which would cause the program to get the real-world time instead of the simulated time. We now patch the VDSO itself at program start to reliably intercept these functions in such cases.
Implemented basic signal emulation. This allows managed code to install signal handlers and send and receive signals within the simulation. (Sending a signal from outside the simulation to a managed process is still not supported). Notably this support is required to handle golang programs, and allows simulated processes to be shut down cleanly by scheduling
killprocesses to send appropriate signals.
- Added an experimental option for strace-style logging of syscalls made by managed code.
--progressoption, which periodically updates stderr with the simulation progress.
- Nightly shadow benchmarks are now being run to help detect performance regressions. We also have the ability to run the benchmark on our own development branches to investigate performance changes before merging these branches into Shadow. Benchmark results for a 5% Tor network are published publicly, but are only intended to be useful for Shadow developers.
- Added a library for overriding crypto functions. When enabled, this option overrides some openssl APIs with “no-op” implementations. This is a reimplementation of a feature previously available in Shadow’s “tor plugin”. It is intended primarily to improve the performance of tor simulations on hardware without accelerated AES operations, and is only supported on Debian 11.
The Shadow team
We have released Shadow 2.0! Give it a try, and let us know if you run into any issues!
- Add unnamed unix sockets created with socketpair()
- Add support for the ‘O_DIRECT’ flag to pipes (packet mode)
- poll and epoll: wait for timeout even with no fd’s
- Set default min runahead to 1ms. This fixes a performance regression in many-host simulations using multiple CPUs.
We have been collaborating with The Tor Project to use Shadow in its development and testing.
We have been running Shadow simulations inside a Gitlab CI pipeline to help develop and tune improved congestion control algorithms in the upcoming 0.4.7 release. This is the first major application of Shadow inside the Tor Project itself, and we plan to use the pipeline we’ve developed for further testing and profiling.
We have also been making progress on running Arti (the experimental new Rust implemenation of tor) under Shadow: . In the course of this work we have fixed some bugs in Arti, fixed an upstream bug in the crate, and found and identified and fixed several emulation accuracy bugs in Shadow.
The Shadow team