I need to run a specific core a hard real time application, I cannot tolerate interruptions higher than 1 usec. Unfortunately, when a CAL/TLB interrupt kicks in, the interruption can reach up to 6-7 usecs.
Is it possible to disable all interrupts on a single specific core? I know that stopping TLB flushes can cause problems but my application does setup memory at startup and virt to phys mappings are static.
I know linux has spinlock_irq_disable but how do I assure that this API is called on the core I need it?
Is not possible to do it from normal application that has root rights?
Related
it is easy to allocate a process to particular core , but how to make sure only that process should run on that particular core or cores.Rest processes can run on other cores. Please help me in this.
i got answer from similar question
Add the parameter isolcpus=[cpu_number] to the Linux kernel command line form the boot loader during boot. This will instruct the Linux scheduler to no run any regular tasks on that CPU unless specifically requested using cpu affinity.
Use IRQ affinity to set other CPUs to handle all interrupts so that your isolated CPU will not received any interrupt.
Use CPU affinity to fix your specific task to the isolated CPU
Is it possible (...knowing full well that this is crazy and seriously ill-advised...) to have a J2EE application running in a Java app server (using weblogic presently), and have a native executable process started, used, and stopped as part of this Java application's lifecycle? (Note: this is not JNI, it's actually a separate native process. It's unix/linux, but should also run on windows.) I haven't found any docs on the subject -- and for good reason, probably.
Background: The native process is actually some monolithic 3rd party software package that is un-hackable and there's no API other than stdin/stdout. The Java app requires the native app to perform certain services. I can easily wrap the native process via ProcessBuilder and start/stop and communicate with it (using stdin/stdout). For testing purposes I have a simple exe (C++) that communicates via stdin/stdout and can receive "start", "shutdown" and performs a simple "echo" service. (The "start" is a no-op, but simply returns "ok" if the native process started successfully.)
So, ideally, when the app server is started/shutdown, and/or the deployed Java app is started/shutdown, the associated native process can also be started/shutdown. And ideally, this can happen cleanly & reliably (no lingering processes after shutdown, all startup failures logged, the lifecycle timing issues synchronized).
If this actually worked, then "part 2" of the question would be if this could actually work in a cluster/failover environment. The native process could be tied to a platform and software-specific monitoring & management service, but I'd like to have everything bundled and managed with the Java app, if possible.
If Glassfish or any other OSGi type environment would make this simpler, please feel free to let me know (it could be an option... I'd prefer Glassfish, but WLS is the blanket mandate.)
I'm trying to put together a proof-of-concept, but any clear answer "yes, I've done it" or "no, it won't work" would be much appreciated & a huge time-saver (with supporting doc links, if you have them).
Edit: just to clarify (the subject may be misleading): there is a considerable Java application running as well (which I've written & can freely modify as necessary); the 3rd party native process just performs a service that the Java application requires. I'm not merely trying to manage a native process via an app server.
The answer to part 1 is yes, it is absolutely possible to have a Java application server manage a native system process. It sounds like you've pretty much figured this out for yourself, if you're thinking about using a ProcessBuilder to spawn the external program and interact with it. That's pretty much the way to do it.
I have used exactly that kind of setup in the past to implement a media transcoding service on top of a Java server (the Java server spawned transcoding jobs via ffmpeg processes, monitoring their status and reporting back to the rest of the application on success/failure/etc.). How cleanly it can all be done depends upon how you implement it and upon the behavior of your external app (i.e. is it guaranteed to respond gracefully and quickly to a shutdown request?), but it will be very difficult (if not impossible) to get it completely perfect. At a minimum, if someone does a kill -9 on your Java server process, there is no way for you to gracefully shut down the native process, at least not until the server is restarted and you see that the native process is already running.
The second part depends upon exactly what you mean by "work in a cluster/failover environment". In terms of managing the native process, if you can start it and interact with it in Java then you can also manage it in Java. But if you mean you want perfect failover behavior such that if the node with the native process on it goes down then a new node automatically resumes the process in the exact same state as it was before, then that may be very difficult or even impossible. But, if you abstract out interactions with the external process so that it just appears as a service that your Java code interacts with (for instance, perhaps by sending requests to some facade class that understands how to interact with and manage the external process) then you should be able to get some fairly good results.
The transcoding service that I implemented ran in a clustered environment (using JBoss/Tomcat), and the way it worked was that when a transcoding job was requested a message would be dispatched. This message would be received by a coordinating class that would manage the queue of transcode requests, spawning jobs as worker processes became available. The state of the queue was replicated across the cluster, so if the node running the ffmpeg processes went down the currently scheduled jobs would be remembered, and then resumed as soon as a suitable node was available again (the transcoding service was configurable so that it could be enabled/disabled per node). In practice the system proved to be quite robust.
My Grails application is running in a development environment. I still didn't go into production, but in any case, is it normal that my Grails application is requiring 230 MB at startup only (with an empty bootstrap and no request handled so far)?
Do you know why this is the case, how to improve memory usage in development mode and, most important, whether it is reduced in production environment?
To answer your questions, yes - it is normal. It's especially normal if you have a lot GSPs in your application. GSPs are runtime compiled so you can speed up their generation by increasing your permgen space.
You can improve memory use and performance in general by making sure that you are passing the '-server' flag when you load your server JVM.
I wouldn't blame all that memory usage just on Grails. Because it uses an embedded Tomcat (Jetty in older versions) there will be a decent amount of overhead even when running an empty application.
IMO, 230MB is a lot of memory use for a Java application. High memory usage is just part of life when writing jvm based applications.
My online Grails applications run in a VPS with only 512MB (which includes a Drupal CMS, Apache, the email services, ... and the Tomcat to run GRails) so you can definitely tune your application to use less memory
I would like to start a new instance of a wcf service host from another (UI) application. I need the service to be out of process because I want to make use of the entire 1.4GB memory limit for a 32bit .NET process.
The obvious method is to use System.Diagnostics.Process.Start(processStartInfo) but I would like to find out whether it is a good way or not. I am planning on bundling the service host exe with the UI application. When I start the process, I will pass in key parameters for the WCF service (like ports and addresses etc). The UI application (or other applications) will then connect to this new process to interact with the service. Once the service has no activity for a while, it will shut itself down or the UI can explicitly make a call to shut the service down.
You can definitely do this:
create a console app which hosts your ServiceHost
make that console app aware of a bunch of command line parameters (or configure them in the console app's app.config)
launch the console app using Process.Start() from your UI app
That should be fairly easy to do, I'd say.
Perhaps I'm completely offbase here, but I don't think there is a 1.4 GB memory limit for .NET processes. The memory allocated for each process is managed by the OS. For 32-bit opeating systems, there is a 4 GB memory space available, but that is shared among all of the processes. So while it may appear that there is only 1.4 GB available, it's not technically true.
The only reason I bring that up is to say that the other way to approach this would be to load your WCF service inside a separate AppDomain within your UI application. The System.AppDomain class can be thought of as a lightweight process within a process. AppDomains can also be unloaded when you are finished with them. And since WCF can cross AppDomain boundaries as well as process boundaries, it's simply another consideration.
If you are not familiar with AppDomains, the approach that #marc_s recommended is the most straightforward. However, if you are looking for an excuse to learn about AppDomains, this would be a great opportunity to do so.
I have several different c# worker applications that run various continuous tasks: sending emails from queue, importing new orders from website database to orders database, making database backups and restores, running data processing for OLTP -> OLAP, and other related tasks. Before, I released these as windows services, but currently I release them as regular console applications. They are all based on a common task runner framework I created, and I am happy with that, however I am not sure what is the best way to deploy these types of applications. I like the console version because it is quick and easy, and it is possible to quickly see program activity and output. The downside is that the worker computer has several console screens running and it gets messy. On the other hand the service method seems to take to long to deploy and I have to go through event logs to see messages. What are some experiences/comments on this?
I like the console app approach. I typically have things set up so I can pass a switch like -unattended that suppresses the console screen.
Windows Service would be a good choice, it runs in the background no matter if you close current session, also you can configure it to start automatically after windows restart when performing a patches update on the server. You can log important messages to event viewer or database table.
For a thing like this, the standard way of doing it is with Windows services. You want the service to run on the network account so it won't require a logged in user.
I worked on something a few years ago that had similar issues. Logically I needed a service, but sometimes I needed to see what was going on and generally I wanted a history. So I developed a service which did the work, any time it wanted to log, it called to it's subscribers (implemented as an observer pattern).
The service registered it's own data logger (writing to a database) and at run time, the user could run a GUI which connected to the service using remoting to become a live listener!
I'm going to vote for Windows Services. It's going to get to be a real pain managing those console applications.
Windows Service deployment is easy: after the initial install, you just turn them off and do an XCOPY. No need to run any complicated installers. It's only semi-complicated the first time, and even then it's just
installutil MyApp.exe
Configre the services to run under a domain account for the best security and easiest interop with other machines.
Use a combination of event logs (with Error, Warning, and Information) for important notifications, and just dump verbose logging to a text file.
Why not get the best of all worlds and use something like:
http://topshelf-project.com/
It will allow you to run your program as command line or a windows service.
I'm not sure if this applies to your applications or not, but when I have some console applications that are not dependent on user input or they are the kind of applications that just do their job and quit, I run such programs on a virtual server, this way I don't see a screen popping up when I'm working, and virtual servers are easy to create and restart.
We regularly use windows services as the background processes. I don't like command-line apps as you need to be logged into the server for them to run. Services run in the background all the time (assuming they're auto-start). They're also trivial to install w/the sc.exe command-line tool that's in windows. I like it better than the bloat-ware that is installutil.exe. Of course installutil does more, but I don't need what it does. I just want to register my service.
We've also created a infrastructure where we have a generic service .exe that loads .DLLs based on an interface definition, so adding a new "service" is as simple as dropping in a new DLL and restarting the service host.
However, we started to move away from services. The problem we have with them is that they lock up the DLLs (for obvious reasons) so it's a pain to upgrade them. We need to stop, upgrade and then restart. Not hard, but additional steps. Instead we're moving to special "pages" in our asp.net apps that run the actual background jobs we need done. There's still a service, but all it does it invoke the asp.net pages so it doesn't lock up any of our DLLs. Then we can replace the DLLs in the asp.net bin directory and normal asp.net rules for app-domain restart kick in.