Hacker News Reader

Comment by gwd | Hacker News Reader

Yes, at some point when I was a student I went out to VMWare and gave a presentation to a team of their engineers.

The thing about execution replay is that, in the general case, it's incredibly fragile. I happened to chat to some people from VMWare in 2008 and they said that there was a team of 10 people whose only job it was to fix execution replay when it broke. The main thing I learned from my PhD on execution replay was how to debug insane bugs.

Presumably rr, focusing on processes, offers a more constrained environment that's easier to log & replay.

View on HN

> Presumably rr, focusing on processes, offers a more constrained environment that's easier to log & replay.

As someone who's worked elsewhere on time travel debug, I'm really curious on @roca's take on this - because I'd have expected a full-VM solution to be easier to make reliable.

Hardware-level behaviour sounds harder but it's well-constrained. The behaviour an OS can rely on from the hardware is large but well-documented and slow to change.

In contrast, the process boundary is really ill-defined and permeable. Also, when you need things to be precisely the same, you notice bits of kernel behaviour leaking into the user space ABI in unexpected ways.

My instinct is the same as yours, but I'm also well aware of the "grass is greener on the other side of the fence" cognitive bias :-). So without actually having written a record-and-replay hypervisor, I hesitate to judge.

If you are doing a single core virtual machine, the hypervisor is properly implemented, and you can change the hypervisor then it is fairly straightforward. You just funnel everything through a single guest memory write function and then instrument that.

Unfortunately, virtual machines are basically only ever used in multicore configurations which are fundamentally unconstrained multithreading which makes record-replay based (versus full trace based) implementations impossible (or at least very challenging) unless you serialize.

The device driver interface is most likely going to be trapping accesses and communicating back via asynchronous shared memory writes if not pass-through or paravirtualized. I suspect most hypervisor implementations will not even allow you to intercept/trace those operations externally, and if they do it is likely going to be hard to sequence them relative to VM execution/instruction stream as they are rooted in a shared memory interface.

If they are paravirtualized, or just in general, the hypercall interface is also usually very ad-hoc and ill-defined and also likely not possible to intercept or trace. The scope is probably going to be more like trying to represent the effects of every random vendor-specific ioctl.

The assumption here is that you're modifying the hypervisor itself. No-one expects to be able to bolt a record-replay tool onto a third-party hypervisor.

Well, that is what you did with rr (bolted it onto a third-party unmodified OS) so I thought I would cover all the bases.

Once you are in modification land it really depends on the kind of hypervisor and its implementation. Like doing it for a OS, a well-designed, consistent, and tight interface makes it pretty easy. But, the unconstrained multithreading for multicore VMs makes it problematic for record-replay solutions if you do not want to serialize.

Yes, this is true. Genuine parallelism will absolutely be a problem there - similarly to how it is for user-level record/replay tech. And serialising is going to have more impact when you're recording a whole OS vs just one application.

Microsoft's TTD (a user-level tech) does allow genuine parallelism, I believe. But my understanding is that it has slower (though still very impressive) single-threaded performance as a result, so there's a tradeoff. But perhaps you could do something more like that.

We believe it's possible to handle parallelism in record/replay tech while still having good single-threaded performance - but it implies a much more complex implementation to do so safely.

The literature on Microsoft TTD [1][2] indicates it is a instruction emulator full-trace approach. It is a little unclear if it is a basic emulator or a instrumented JIT approach, but the reported ~10x overhead is more consistent with a basic emulator. Though maybe their recording implementation is just inefficient causing such excess overhead.

[1] https://msrc.microsoft.com/blog/2019/05/time-travel-debuggin...

[2] https://www.usenix.org/legacy/events/vee06/full_papers/p154-...

The Linux kernel supports pretty powerful APIs for inspecting and managing processes, like ptrace and seccomp and /proc, from another userspace process. There is no comparable set of APIs for hypervisors, unless Xen has something maybe?

And we don't handle unrestricted multithreading in rr. If we wanted to it would complicate things a lot and I don't know if existing kernel APIs would be adequate. We're skating along the edge of feasibility as it is.

The interface between userspace processes and the kernel and CPU is pretty complicated. It's hard for me to tell whether it's more or less complicated than the interface between guests and the hypervisor (given most record and replay scenarios don't require virtualizing exotic hardware).

It probably is easier to debug rr replay bugs than VM-level replay bugs, partly because rr is lighter-weight. rr replay bugs are still insane, but it's all relative :-).

My first prototype of execution replay I implemented was for Linux processes (in like, 2000 or something), which I did by modifying Linux. The primary thing I did there was to instrument copy_to_user().

The final version of execution replay that I did was for Xen, which had the asynchronous device drivers. In that case, the goal of the drivers was to have the device do DMA directly into the guest's memory space, and thus had to be modeled as an external processor. If you were content to modify your virtual devices to always doing copying, it would be significantly easier.

But yeah, in both cases you just have so much that the thing underneath might be doing, it's hard to tell which is going to be more fragile. :-)

The final version also did multi-processor VMs, using the paging hardware to track sharing across virtual cpus [1]. The key take-away from that was, "It's slow, but often not as slow as you might think."

[1] https://www.semanticscholar.org/paper/Execution-replay-of-mu...

Reply on news.ycombinator.com