I'm looking for a little advice on "hacking" Mono (and in fact, .NET too).
Context: As part of the Isis2 library (Isis2.codeplex.com) I want to support very fast "zero copy" replication of memory-mapped files on machines that have the right sort of hardware (Infiband NICs), and minimal copying for more standard Ethernet with UDP. So the setup is this: We have a set of processes {A,B....} all linked to Isis2, and some member, maybe A, has a big memory-mapped file, call it F, and asks Isis2 to please rereplicate F onto B, D, G, and X. The library will do this very efficiently and very rapidly, even with heavy use by many concurrent initiators. The idea would be to offer this to HPC and cloud developers who are running big-data applications.
Now, Isis2 is coded in C# on .NET and cross-compiles to Linux via Mono. Both .NET and Mono are managed, so neither wants to let me do zero-copy network I/O -- the normal model would be "copy your data into a managed byte[] object, then use SendTo or SendAsync to send. To receive, same deal: Receive or ReceiveAsync into a byte[] object, then copy to the target location in the file." This will be slower than what the hardware can sustain.
Turns out that on .NET I can hack around the normal memory protections. I built my own mapped file wrapper (in fact based on one posted years ago by a researcher at Columbia). I pull in the Win32Kernel.dll library, and then use Win32 methods to map my file, initiate the socket Send and Receive calls, etc. With a bit of hacking I can mimic .NET asynchronous I/O this way, and I end up with something fairly clean and coded entirely in C# with nothing .NET even recognizes as unsafe code. I get to treat my mapped file as a big unmanaged byte array, avoiding all that unneeded copying. Obviously I'll protect all of this from my Isis2 users; they won't know.
Now we get to the crux of my question: on Linux, I obviously can't load the Win32 kernel dll since it doesn't exist. So I need to implement some basic functionality using core Linux O/S calls: the fmap() call will map my file. Linux has its own form of asynchronous I/O too: for Infiniband, I'll use the Verbs library from Mellanox, and for UDP, I'll work with raw IP sends and signals ("interrupts") on completion. Ugly, but I can get this to work, I think. Again, I'll then try to wrap all this to look as much like standard asynchronous Windows async I/O as possible, for code cleanness in Isis2 itself, and I'll hide the whole unmanaged, unsafe mess from end users.
Since I'll be sending a gigabyte or so at a time, in chunks, one key goal is that data sent in order would ideally be received in the order I post my async receives. Obviously I do have to worry about unreliable communication (causes stuff to end up dropped, and I'll then have to copy). But if nothing is dropped I want the n'th chunk I send to end up in the n'th receive region...
So here's my question: Has anyone already done this? Does anyone have any tips on how Mono implements the asynchronous I/O calls that .NET uses so heavily? I should presumably do it the same way. And does anyone have any advice on how to do this with minimal pain?
One more question: Win32 is limited to 2Gb of mapped files. Cloud systems would often run Win64. Any suggestions on how to maximize interoperability while allowing full use of Win64 for those who are running that? (A kind of O/S reflection issue...)
Related
in my project i injected a DLL(64-bit Windows 10) in to a external process with Manual-map & Thread-hijacking and i do some stuff in there.
In current state i use "RtlCreateUserThread" to create a new thread and do some extra workload in there to distribute it for better performance.
My question is now... Is it possible to access other threads from the current process (hijack it) and add your own workload/code there. Without creating a new thread?
I didn't found anything helpful yet in the internet and the code i used and modified for Thread-hijacking seems to only work for a DLL file. Because i am pretty new to C++ i am still learning i am already thankful for any help.
(If you want to see the source for injector Google GHInjector your find the library on github.)
It is possible, but so complicated and may not work in all cases.
You need to splice existing thread's machine codes, so you will need write access to code page memory.
Logic:
find thread id and thread handle, then suspend thread with SuspendThread WINAPI call
suspended thread can be in wait state or in system DLL call now, so you need to analyze current execution stack, backtrace it and find execution address from application space. You need API functions StackWalk, and PDB files in some cases. Also it depends on running architecture (x86, amd64, ...). Walk through stack until your EIP/RIP will not be in application memory address space
decode machine instruction (it will be 'call') and splice next instructions to your function call. You need to use __declspec(naked) declared function or ASM implemented one for execute your code and replaced instructions.
ResumeThread
This method may work only once because no guarantees that application code is executed in loop.
I am evaluating different multiprocessing libraries for a fault tolerant application. I basically need any process to be allowed to crash without stopping the whole application.
I can do it using the fork() system call. The limit here is that the process can be created on the same machine, only.
Can I do the same with MPI? If a process created with MPI crashes, can the parent process keep running and eventually create a new process?
Is there any alternative (possibly multiplatform and open source) library to get the same result?
As reported here, MPI 4.0 will have support for fault tolerance.
If you want collectives, you're going to have to wait for MPI-3.something (as High Performance Mark and Hristo Illev suggest)
If you can live with point-to-point, and you are a patient person willing to raise a bunch of bug reports against your MPI implementation, you can try the following:
disable the default MPI error handler
carefully check every single return code from your MPI programs
keep track in your application which ranks are up and which are down. Oh, and when they go down they can never get back. but you're unable to use collectives anyway (see my opening statement), so that's not a huge deal, right?
Here's an old paper (back when Bill still worked at Argonne. I think it's from 2003):
http://www.mcs.anl.gov/~lusk/papers/fault-tolerance.pdf . It lays out the kinds of fault tolerant things one can do in MPI. Perhaps such a "constrained MPI" might still work for your needs.
If you're willing to go for something research quality, there's two implementations of a potential fault tolerance chapter for a future version of MPI (MPI-4?). The proposal is called User Level Failure Mitigation. There's an experimental version in MPICH 3.2a2 and a branch of Open MPI that also provides the interfaces. Both are far from production quality, but you're welcome to try them out. Just know that since this isn't in the MPI Standard, the function prefixes are not MPI_*. For MPICH, they're MPIX_*, for the Open MPI branch, they're OMPI_* (though I believe they'll be changing theirs to be MPIX_* soon as well.
As Rob Latham mentioned, there will be lots of work you'll need to do within your app to handle failures, though you don't necessarily have to check all of your return codes. You can/should use MPI error handlers as a callback function to simplify things. There's information/examples in the spec available along with the Open MPI branch.
I have 2 processes, and I would like one of the processes to talk to the other one with a high data throughput. I have tried IPC(boost::iterprocess specifically) and sockets but their performance/throughput is too slow to use.
My fallback option is to launch the 2nd process as an attached child of the first(load its dll, create the 'tool', etc), which has the best performance, as they are technically the same process at that point, and passing data is just calling the interface functions with the DLL.
I'm looking for ways to avoid doing it like this but still have that degree of performance. Is it possible to set up a DLL that 2 processes can load and somehow share memory space between? Are IPC and sockets the only options here?
On windows, you can use named pipes. They were once considered more efficient than sockets when used locally. They've gone out of vogue however. You can learn more here microsoft docs on named pipes
What is "IPC" in your question? Sockets, pipes, shared memory are all ways to do IPC. And yes, you can use shared memory on Windows, Linux and other general-purpose systems. In C++ you can just declare a memory block as shared (at least on Windows) or you can call Memory Mapped File (MMF) functions. On Linux and BSD you use Memory Mapped File functions as well.
MMFs are the fastest possible way besides transforming a second process into the DLL. Named pipes and anything else is slower.
For local processes you can use a shared file. If you memory map the file it would be much faster.
I'm working on a project which will entail multiple devices, each with an embedded (ARM) processor, communicating. One development approach which I have found useful in the past with projects that only entailed a single embedded processor was develop the code using Visual Studio, divided into three portions:
Main application code (in unmanaged C/C++ [see note])
I/O-simulating code (C/C++) that runs under Visual Studio
Embedded I/O code (C), which Visual Studio is instructed not to build, runs on the target system. Previously this code was for the PIC; for most future projects I'm migrating to the ARM.
Feeding the embedded compiler/linker the code from parts 1 and 3 yields a hex file that can run on the target system. Running parts 1 and 2 together yields code which can run on the PC, with the benefit of better debugging tools and more precise control over I/O behavior (e.g. I can make the simulation code introduce certain types of random hiccups more easily than I can induce controlled hiccups on real hardware).
Target code is written in C, but the simulation environment uses C++ so as to simulate I/O registers. For example, I have a PortArray data structure; the header file for the embedded compiler includes a line like unsigned char LATA # 0xF89; and my header file for simulation includes #define LATA _IOBIT(f89,1) which in turn invokes a macro that accesses a suitable property of an I/O object, so a statement like LATA |= 4; will read the simulated latch, "or" the read value with 4, and write the new value. To make this work, the target code has to compile under C++ as well as under C, but this mostly isn't a problem. The biggest annoyance is probably with enum types (which behave as integers in C, but have to be coaxed to do so in C++).
Previously, I've used two approaches to making the simulation interactive:
Compile and link a DLL with target-application and simulation code, and have VB code in the same project which interacts with it.
Compile the target-application code and some simulation code to an EXE with instance of Visual Studio, and use a second instance of Visual Studio for the simulation-UI. Have the two programs communicate via TCP, so nearly all "real" I/O logic is in the simulation program. For example, the aforementioned `LATA |= 4;` would send a "read port 0xF89" command to the TCP port, get the response, process the received value, and send a "write port 0xF89" command with the result.
I've found the latter approach to run a tiny bit slower than the former in some cases, but it seems much more convenient for debugging, since I can suspend execution of the unmanaged simulation code while the simulation UI remains responsive. Indeed, for simulating a single target device at a time, I think the latter approach works extremely well. My question is how I should best go about simulating a plurality of target devices (e.g. 16 of them).
The difficulty I have is figuring out how to make each simulated instance get its own set of global variables. If I were to compile to an EXE and run one instance of the EXE for each simulated target device, that would work, but I don't know any practical way to maintain debugger support while doing that. Another approach would be to arrange the target code so that everything would compile as one module joined together via #include. For simulation purposes, everything could then be wrapped into a single C++ class, with global variables turning into class-instance variables. That would be a bit more object-oriented, but I really don't like the idea of forcing all the application code to live in one compiled and linked module.
What would perhaps be ideal would be if the code could load multiple instances of the DLL, each with its own set of global variables. I have no idea how to do that, however, nor do I know how to make things interact with the debugger. I don't think it's really necessary that all simulated target devices actually execute code simultaneously; it would be perfectly acceptable for simulation instances to use cooperative multitasking. If there were some way of finding out what range of memory holds the global variables, it might be possible to have the 'task-switch' method swap out all of the global variables used by the previously-running instance and swap in the contents applicable to the instance being switched in. Although I'd know how to do that in an embedded context, though, I'd have no idea how to do that on the PC.
Edit
My questions would be:
Is there any nicer way to allow simulation logic to be paused and examined in VS2010 debugger, while keeping a responsive UI for the simulator front-end, than running the simulator front end and the simulator logic in separate instances of VS2010, if the simulation logic must be written in C and the simulation front end in managed code? For example, is there a way to tell the debugger that when a breakpoint is hit, some or all other threads should be allowed to keep running while the thread that had hit the breakpoint sits paused?
If the bulk of the simulation logic must be source-code compatible with an embedded system written in C (so that the same source files can be compiled and run for simulation purposes under VS2010, and then compiled by the embedded-systems compiler for use in real hardware), is there any way to have the VS2010 debugger interact with multiple simulated instances of the embedded device? Assume performance is not likely to be an issue, but the number of instances will be large enough that creating a separate project for each instance would be likely be annoying in the absence of any way to automate the process. I can think of three somewhat-workable approaches, but don't know how to make any of them work really nicely. There's also an approach which would be better if it's possible, but I don't know how to make it work.
Wrap all the simulation code within a single C++ class, such that what would be global variables in the target system become class members. I'm leaning toward this approach, but it would seem to require everything to be compiled as a single module, which would annoyingly affect the design of the target system code. Is there any nice way to have code access class instance members as though they were globals, without requiring all functions using such instances to be members of the same module?
Compile a separate DLL for each simulated instance (so that e.g. if I want to run up to 16 instances, I would include 16 DLL's in the project, all sharing the same source files). This could work, but every change to the project configuration would have to be repeated 16 times. Really ugly.
Compile the simulation logic to an EXE, and run an appropriate number of instances of that EXE. This could work, but I don't know of any convenient way to do things like set a breakpoint common to all instances. Is it possible to have multiple running instances of an EXE attached to a single debugger instance?
Load multiple instances of a DLL in such a way that each instance gets its own global variables, while still being accessible in the debugger. This would be nicest if it were possible, but I don't know any way to do so. Is it possible? How? I've never used AppDomains, but my intuition would suggest that might be useful here.
If I use one VS2010 instance for the front-end, and another for the simulation logic, is there any way to arrange things so that starting code in one will automatically launch the code in the other?
I'm not particularly committed to any single simulation approach; while it might be nice to know if there's some way of slightly improving the above, I'd also like to know of any other alternative approaches that could work even better.
I would think that you'd still have to run 16 copies of your main application code, but that your TCP-based I/O simulator could keep a different set of registers/state for each TCP connection that comes in.
Instead of a bunch of global variables, put them into a single structure that encompasses the I/O state of a single device. Either spawn off a new thread for each socket, or just keep a list of active sockets and dedicate a single instance of the state structure for each socket.
the simulators I have seen that handle multiple instances of the instruction set/processor are designed that way. There is a structure usually that contains a complete set of registers, and a new pointer or an array of these structures are used to multiply them into multiple instances of the processor.
I need a way to store large data chunks(~1-2MB) at a high rate (~200-300Mbit/s).
After some research I found several options:
aio_write
Direct_IO
Carbon File Manager's PBWriteForkAsync()
default fwrite(), wrapped in a block and dispatched via GCD
NSData's appendData in an NSOperation
...
This wiki page describes the state of aio_write under Linux. What I didn't find was a similar page about the state of aio_write for Mac OS X.
NSOperation or Blocks+GCD seems to be a technique to achieve non-blocking IO. It is used in several open source IO libraries (e.g. https://github.com/mikeash/MAAsyncIO)
Has someone with a similar problem found a suitable solution?
Currently I tend towards PBWriteForkAsync as it takes some 'tuning'parameters. It also should be 64-bit safe.
I don't know MacOS very well, but I'd also try open and write syscalls from unistd.h with the non-blocking option O_NONBLOCK. reference
You should use unbuffered I/O for the writes, in Carbon this is FSWriteFork() with kFSNoCacheBit, in BSD use fcntl() with F_NOCACHE.
Rather than use the system's non-blocking IO, you may want to consider a worker thread to write the blocks sequentially using a queue. This will give you more control and may end up being simpler, particularly if you want to monitor the queue to see if you are keeping up or not.
See here for more information.