I am using grpc for bidirection streaming service. Beause this is an online real-time service, and we don't want our clients to wait too long, so we want to limit the rpcs to a certain number and reject the extra requests.
We are able to do it in Python with following code:
grpcServer = grpc.server(futures.ThreadPoolExecutor(max_workers=10), maximum_concurrent_rpcs=10)
In such case, only 10 requests will be processed, and other request will be rejected, the clients will throw a RESOURCE_EXHAUSTED error.
However, we find it hard to do it in C++. We use code following grpc sync server limit handle thread
ServerBuilder builder;
builder.SetSyncServerOption(ServerBuilder::SyncServerOption::NUM_CQS, 10);
grpc::ResourceQuota quota;
quota.SetMaxThreads(10 * 2);
builder.SetResourceQuota(quota);
We are using grpc in many services. Some using kaldi, some using libtorch.
In some cases, the above code behaves normally, it processes 10 requests at a time, and rejects other requests. and the processing speed (our service requires a log of cpu calculation) is ok.
In some cases, it only accepts 9 request at a time.
In some cases, it accepts 10 request at a time, but the processing speed is significantly lower than before.
We also tried
builder.AddChannelArgument(GRPC_ARG_MAX_CONCURRENT_STREAMS, 10);
But it is useless becuase GRPC_ARG_MAX_CONCURRENT_STREAMS is the max_concurrent_rpcs on a single http connection.
Could someone please point out the equivalent C++ version of the Python code? We want to make our service to handle 10 requests at a time, and rejects other request, we do not want any service to wait in the queue.
Related
I am trying to switch my application to the async version from apache http-components client . The goal is to be able to handle more outbound connections (in the near future).
The payload of the requests is quite small (<5KB)
The endpoints I hit are around 20 in number.
With sync version of apache http client, the through put is about 200 requests/sec.
The average response time is about 100ms/request.
I abort the requests after a max of 180ms.
After switching to Async, the response time went up by 20ms/request.
The throughput also reduced to 160/sec. The number of aborted requests doubled.
This is after fine tuning the application a lot.
Is there anything I can do to improve the performance of async client?
I set maxConnectionsPerRoute high. Have a large Connection pool.
Are there any params that are key to getting the most out of async client?
Did you forget to set maxConnTotal?
The default maxConnTotal is 20, this is a global limit.
I forgot to set it once.
I think I know what is happening here, but would appreciate a confirmation and/or reading material that can turn that "think" into just "know", actual questions at the end of post in Tl,DR section:
Scenario:
I am in the middle of testing my MVC application for a case where one of the internal components is stalling (timeouts on connections to our database).
On one of my web pages there is a Jquery datatable which queries for an update via ajax every half a second - my current task is to display correct error if that data requests times out. So to test, I made a stored procedure that asks DB server to wait 3 seconds before responding, which is longer than the configured timeout settings - so this guarantees a time out exception for me to trap.
I am testing in Chrome browser, one client. Application is being debugged in VS2013 IIS Express
Problem:
Did not expect the following symptoms to show up when my purposeful slow down is activated:
1) After launching the page with the rigged datatable, application slowed down in handling of all requests from the client browser - there are 3 other components that send ajax update requests parallel to the one I purposefully broke, and this same slow down also applied to any actions I made in the web application that would generate a request (like navigating to other pages). The browser's debugger showed the requests were being sent on time, but the corresponding break points on the server side were getting hit much later (delays of over 10 seconds to even a several minutes)
2) My server kept processing requests even after I close the tab with the application. I closed the browser, I made sure that the chrome.exe process is terminated, but breakpoints on various Controller actions were still getting hit for 20 minutes afterward - mostly on the actions that were "triggered" by automatically looping ajax requests from several pages I was trying to visit during my tests. Also breakpoints were hit on main pages I was trying to navigate to. On second test I used RawCap monitor the loopback interface to make sure that there was nothing actually making requests still running in the background.
Theory I would like confirmed or denied with an alternate explanation:
So the above scenario was making looped requests at a frequency that the server couldn't handle - the client datatable loop was sending them every .5 seconds, and each one would take at least 3 seconds to generate the timeout. And obviously somewhere in IIS express there has to be a limit of how many concurrent requests it is able to handle...
What was a surprise for me was that I sort of assumed that if that limit (which I also assumed to exist) was reached, then requests would be denied - instead it appears they were queued for an absolutely useless amount of time to be processed later - I mean, under what scenario would it be useful to process a queued web request half an hour later?
So my questions so far are these:
Tl,DR questions:
Does IIS Express (that comes with Visual Studio 2013) have a concurrent connection limit?
If yes :
{
Is this limit configurable somewhere, and if yes, where?
How does IIS express handle situations where that limit is reached - is that handling also configurable somewhere? ( i mean like queueing vs. immediate error like server is busy)
}
If no:
{
How does the server handle scenarios when requests are coming faster than they can be processed and can that handling be configured anywhere?
}
Here - http://www.iis.net/learn/install/installing-iis-7/iis-features-and-vista-editions
I found that IIS7 at least allowed unlimited number of silmulatneous connections, but how does that actually work if the server is just not fast enough to process all requests? Can a limit be configured anywhere, as well as handling of that limit being reached?
Would appreciate any links to online reading material on the above.
First, here's a brief web server 101. Production-class web servers are multithreaded, and roughly one thread = one request. You'll typically see some sort of setting for your web server called its "max requests", and this, again, roughly corresponds to how many threads it can spawn. Each thread has overhead in terms of CPU and RAM, so there's a very real upward limit to how many a web server can spawn given the resources the machine it's running on has.
When a web server reaches this limit, it does not start denying requests, but rather queues requests to handled once threads free up. For example, if a web server has a max requests of 1000 (typical) and it suddenly gets bombarded with 1500 requests. The first 1000 will be handled immediately and the further 500 will be queued until some of the initial requests have been responded to, freeing up threads and allowing some of the queued requests to be processed.
A related topic area here is async, which in the context of a web application, allows threads to be returned to the "pool" when they're in a wait-state. For example, if you were talking to an API, there's a period of waiting, usually due to network latency, between sending the request and getting a response from the API. If you handled this asynchronously, then during that period, the thread could be returned to the pool to handle other requests (like those 500 queued up requests from the previous example). When the API finally responded, a thread would be returned to finish processing the request. Async allows the server to handle resources more efficiently by using threads that otherwise would be idle to handle new requests.
Then, there's the concept of client-server. In protocols like HTTP, the client makes a request and the server responds to that request. However, there's no persistent connection between the two. (This is somewhat untrue as of HTTP 1.1. Connections between the client and server are sometimes persisted, but this is only to allow faster future requests/responses, as the time it takes to initiate the connection is not a factor. However, there's no real persistent communication about the status of the client/server still in this scenario). The main point here is that if a client, like a web browser, sends a request to the server, and then the client is closed (such as closing the tab in the browser), that fact is not communicated to the server. All the server knows is that it received a request and must respond, and respond it will, even though there's technically nothing on the other end to receive it, any more. In other words, just because the browser tab has been closed, doesn't mean that the server will just stop processing the request and move on.
Then there's timeouts. Both clients and servers will have some timeout value they'll abide by. The distributed nature of the Internet (enabled by protocols like TCP/IP and HTTP), means that nodes in the network are assumed to be transient. There's no persistent connection (aside from the same note above) and network interruptions could occur between the client making a request and the server responding to the request. If the client/server did not plan for this, they could simply sit there forever waiting. However, these timeouts are can vary widely. A server will usually timeout in responding to a request within 30 seconds (though it could potentially be set indefinitely). Clients like web browsers tend to be a bit more forgiving, having timeouts of 2 minutes or longer in some cases. When the server hits its timeout, the request will be aborted. Depending on why the timeout occurred the client may receive various error responses. When the client times out, however, there's usually no notification to the server. That means that if the server's timeout is higher than the client's, the server will continue trying to respond, even though the client has already moved on. Closing a browser tab could be considered an immediate client timeout, but again, the server is none the wiser and keeps trying to do its job.
So, what all this boils down is this. First, when doing long-polling (which is what you're doing by submitting an AJAX request repeatedly per some interval of time), you need to build in a cancellation scheme. For example, if the last 5 requests have timed out, you should stop polling at least for some period of time. Even better would be to have the response of one AJAX request initiate the next. So, instead of using something like setInterval, you could use setTimeout and have the AJAX callback initiate it. That way, the requests only continue if the chain is unbroken. If one AJAX request fails, the polling stops immediately. However, in that scenario, you may need some fallback to re-initiate the request chain after some period of time. This prevents bombarding your already failing server endlessly with new requests. Also, there should always be some upward limit of the time polling should continue. If the user leaves the tab open for days, not using it, should you really keep polling the server for all that time?
On the server-side, you can use async with cancellation tokens. This does two things: 1) it gives your server a little more breathing room to handle more requests and 2) it provides a way to unwind the request if some portion of it should time out. More information about that can be found at: http://www.asp.net/mvc/overview/performance/using-asynchronous-methods-in-aspnet-mvc-4#CancelToken
I am implementing a hub/servers MPI application. Each of the servers can get tied up waiting for some data, then they do an MPI Send to the hub. It is relatively simple for me to have the hub waiting around doing a Recv from ANY_SOURCE. The hub can get busy working with the data. What I'm worried about is skipping data from one of the servers. How likely is this scenario:
server 1 and 2 do Send's
hub does Recv and ends up getting data from server 1
while hub busy, server 1 gets more data, does another Send
when hub does its next Recv, it gets the more recent server 1 data rather than the older server2
I don't need a guarantee that the order the Send's occur is the order the ANY_SOURCE processes them (though it would be nice), but if I new in practice it will be close to the order they are sent, I may go with the above. However if it is likely I could skip over data from one of the servers, I need to implement something more complicated. Which I think would be this pattern:
servers each do Send's
hub does an Irecv for each server
hub does a Waitany on all server requests
upon completion of one server request, hub does a Test on all the others
of all the Irecv's that have completed, hub selects the oldest server data (there is timing tag in the server data)
hub communicates with the server it just chose, has it start a new Send, hub a new Irecv
This requires more complex code, and my first effort crashed inside the Waitany call in a way that I'm finding difficult to debug. I am using the Python bindings mpi4py - so I have less control over buffers being used.
It is guaranteed by the MPI standard that the messages are received in the order they are sent (non-overtaking messages). See also this answer to a similar question.
However, there is no guarantee of fairness when receiving from ANY_SOURCE and when there are distinct senders. So yes, it is the responsibility of the programmers to design their own fairness system if the application requires it.
We have a requirement to to support 10k+ users, where every user initiate a request and waits for a response from the server (the response can take as long as 20-30 seconds to arrive). it is only one request from the client, and after a long processing by the server, a response will be transmitted and then the connection will disconnect.
in the background, the server will do some DB search and wait for other background processes to notify on completion before responding to the client.
after doing some research i figured out we will need to use something like the atmosphere framework to support websockets/sse event/long polling along with an asynchronous server like netty (=> nettosphere) or jetty.
As for my experience - mostly Java EE world and Tomcat server.
my questions are:
what will be easier to implement in regard to my experience and our requirement: atmosphere + netty or atmoshphere+jetty? which one can scale better, has an easier learning curve and easier to implement other java technologies?
how do u implement in atmosphere a response that is sent only to the originating client and not broadcast to the rest of the clients? (all the examples i found are broadcast).
how can i implement in netty (or jetty) when using the atmosphere framework our response? i.e., the client send a request, after it is received in the server some background processes are run, and when they finish i need to locate the connection and transmit the response. is that achievable?
Some thoughts:
At 10k+ users, with 20-30 second response latency, you likely hit file descriptor limits if using just 1 network interface. Consider a solution that uses multiple network interfaces.
Your description of your request/response can be handled entirely with standard Servlet 3.0, standard HTTP/1.1, Async request handling, and large timeouts.
If your clients are web browsers, and you don't start sending a response from the server until the 20-30 second window, you might hit browser idle timeouts.
Atmosphere and Cometd do the same things, supporting long duration connections, with connection technique fallbacks, and with logical channel APIs.
I believe the AKKA framework will handle this sort of need. I am looking at using it to handle scaling issues possibly with a RabbitMQ to help off load work to potentially other servers that may be added later to scale as needed.
In order to overcome the (apparent) 4 minute idle connection timeout on the Azure load balancer, it seems necessary to send some data down the pipe to the client every now and again to keep the connection from being regarded as idle.
Our controller is set up as an AsyncController, and it fires several different asynchronous methods on other objects, all of which are set up to use IO Completion Ports. Thus, we return from our method immediately, and when the completion packet is processed, IIS hooks back up to the original request so that we can render our View.
Is there any way to periodically send a few bytes down the wire in this case? In a "classic" situation, we could have executed the method and then just spun while we waited, sending data every few seconds until the asynchronous method was complete. But, in this situation, the IIS thread is freed to go do other business, and we hook back up to it in our completion callback. What to do? Is this possible?
While your particular case concerns Windows Azure specific (the 4 minute timeout of LBs), the question is pure IIS / ASP.NET workwise. Anyway, I don't think it is possible to send "ping-backs" to the client while in AsyncController/AsyncPage. This is the whole idea of the AsyncPages/Controllers. The IIS leaves the socket aside having the thread serving other requests. And gets back only when you got the OutstandingOperations to zero with AsyncManager.OutstandingOperations.Decrement(); Only then the control is given back to send final response to the client. And once you are the point of sending response, there is no turning back.
I would rather argue for the architectural approach of why you thing someone would wait 4 minutes to get a response (even with a good animated "please wait")? A lot of things may happen during this time. From browser crash, through internet disruption to total power loss/disruption at client. If you are doing real Azure, why not just send tasks for a Worker Role via a Queue (Azure Storage Queues or Service Bus Queues). The other option that stays in front of you for so long running tasks is to use SingalR and fully AJAXed solution. Where you communicate via SignalR the status of the long running operation.
UPDATE 1 due to comments
In addition to the approach suggested by #knightpfhor this can be also achieved with a Queues. Requestor creates a task with some Unique ID and sends it to "Task submission queue". Then "listens" (or polls at regular/irregular intervals) a "Task completion" queue for a message with given Task ID.
In any way I don't see a reason for keeping client connected for the whole duration of the long running task. There are number of ways to decouple such communication.