grpc-dotnet client: excessive memory usage - grpc

I have a gRPC client integration which recieves messages which are about 65MB in size (date x time tuple arrays mainly). Deserializing the received message seems to allocate about 700MB of extra unmanaged memory on the initial request. Subsequent identical requests do not increase the total process memory consumption anymore.
Is that normal?
Any ideas how to figure out what is going on, or even better, to control this?

It seems this had to do with receiving the data in async contexts. Somehow the task scheduler held on to the Task which was the converted message payload. Adding a Task.Yield() at the end of the async method calling the gRPC client seems to solve it.
Check this discussion:
https://github.com/grpc/grpc-dotnet/issues/1267

Related

handle server shutdown while serving http request

Scenario : The server is in middle of processing a http request and the server shuts down. There are multiple points till where the code has executed. How are such cases typically handled ?. A typical example could be that some downstream http calls had to be made as a part of the incoming http request. How to find whether such calls were made or not made when the shutdown occurred. I assume that its not possible to persist every action in the code flow. Suggestions and views are welcome.
There are two kinds of shutdowns to consider here.
There are graceful shutdowns: when the execution environment politely asks your process to stop (e.g. systemd sends a SIGTERM) and expects it to exit on its own. If your process doesn’t exit within a few seconds, the environment proceeds to kill the process in a more forceful way.
A typical way to handle a graceful shutdown is:
listen for the signal from the environment
when you receive the signal, stop accepting new requests...
...and then wait for all current requests to finish
Exactly how you do this depends on your platform/framework. For instance, Go’s standard net/http library provides a Server.Shutdown method.
In a typical system, most shutdowns will be graceful. For example, when you need to restart your process to deploy a new version of code, you do a graceful shutdown.
There can also be unexpected shutdowns: e.g. when you suddenly lose power or network connectivity (a disconnected server is usually as good as a dead one). Such faults are harder to deal with. There’s an entire body of research dedicated to making distributed systems robust to arbitrary faults. In the simple case, when your server only writes to a single database, you can open a transaction at the beginning of a request and commit it before returning the response. This will guarantee that either all the changes are saved to the database or none of them are. But if you call multiple downstream services as part of one upstream HTTP request, you need to coordinate them, for example, with a saga.
For some applications, it may be OK to ignore unexpected shutdowns and simply deal with any inconsistencies manually if/when they arise. This depends on your application.

Does MPI_Probe return as soon as possible?

Suppose my MPI process is waiting for a very big message, and I am waiting for it with MPI_Probe. Is it correct to suppose the MPI_Probe call will return as soon as the process receives the first notice of the message from the network (like a header with the size or something like)?
I.e., will it return much faster than if I was waiting for the message with MPI_Recv, because it wouldn't need to receive the full message?
The standard is fairly silent on this matter (MPI-3.0, section 3.8.1), but does offer this:
The MPI implementation of MPI_PROBE and MPI_IPROBE needs to guarantee progress:
if a call to MPI_PROBE has been issued by a process, and a send that matches the probe
has been initiated by some process, then the call to MPI_PROBE will return, unless the
message is received by another concurrent receive operation (that is executed by another
thread at the probing process).
Since both MPI_PROBE and MPI_RECV will engage the progress engine, I would doubt there is much difference between the two functions, aside from a memory copy. By engaging the progress engine, it's likely the message will be received (internally) by the MPI implementation. The last step of copying it into the user's buffer can be avoided in MPI_PROBE.
If you are worried about performance, then avoiding MPI_ANY_SOURCE and MPI_ANY_TAG if possible will help most implementations (certainly MPICH) take a faster path.

Call to slow service over HTTP from within message-driven bean (MDB)

I have a message driven bean which serves messages in a following way:
1. It takes data from incoming message.
2. Calls external service via HTTP (literally, sends GET requests using HttpURLConnection), using the data from step 1. No matter how long the call takes - the message MUST NOT be dropped.
3. Uses the outcome from step 2 to persist data (using entity beans).
Rate of incoming messages is:
I. Low most of the time: an order of units / tens in a day.
II. Sometimes high: order of hundreds in a few minutes.
QUESTION:
Having that service in step (2) is relatively slow (20 seconds per request and degrades upon increasing workload), what is the best way to deal with situation II?
WHAT I TRIED:
1. Letting MDB to wait until service is executed, no matter how long it takes. This tends to rollback MDB transactions by timeout and to re-deliver message, increasing workload and making things even worse.
2. Setting timeout for HttpURLConnection gives some guarantees in terms of completion time of MDB onMessage() method, but leaves an open question: how to proceed with 'timed out' messages.
Any ideas are very much appreciated.
Thank you!
In that case you can just increase a transaction timeout for your message driven beans.
This is what I ended up with (mostly, this is application server configuration):
Relatively short (comparing to transaction timeout) timeout for HTTP call. The
rationale: long-running transactions from my experience tend to
have adverse side effects such as threads which are "hung" from app.
server point of view, or extra attention to database configuration,
etc.I chose 80 seconds as timeout value.
Increased up to several minutes re-delivery interval for failed
messages.
Careful adjustment of the number of threads which handle messages
simultaneously. I balanced this value with throughput of HTTP service.

TCP client-server SIGPIPE

I am designing and testing a client server program based on TCP sockets(Internet domain). Currently , I am testing it on my local machine and not able to understand the following about SIGPIPE.
*. SIGPIPE appears quite randomly. Can it be deterministic?
The first tests involved single small(25 characters) send operation from client and corresponding receive at server. The same code, on the same machine runs successfully or not(SIGPIPE) totally out of my control. The failure rate is about 45% of times(quite high). So, can I tune the machine in any way to minimize this.
**. The second round of testing was to send 40000 small(25 characters) messages from the client to the server(1MB of total data) and then the server responding with the total size of data it actually received. The client sends data in a tight loop and there is a SINGLE receive call at the server. It works only for a maximum of 1200 bytes of total data sent and again, there are these non deterministic SIGPIPEs, about 70% times now(really bad).
Can some one suggest some improvement in my design(probably it will be at the server). The requirement is that the client shall be able to send over medium to very high amount of data (again about 25 characters each message) after a single socket connection has been made to the server.
I have a feeling that multiple sends against a single receive will always be lossy and very inefficient. Shall we be combining the messages and sending in one send() operation only. Is that the only way to go?
SIGPIPE is sent when you try to write to an unconnected pipe/socket. Installing a handler for the signal will make send() return an error instead.
signal(SIGPIPE, SIG_IGN);
Alternatively, you can disable SIGPIPE for a socket:
int n = 1;
setsockopt(thesocket, SOL_SOCKET, SO_NOSIGPIPE, &n, sizeof(n));
Also, the data amounts you're mentioning are not very high. Likely there's a bug somewhere that causes your connection to close unexpectedly, giving a SIGPIPE.
SIGPIPE is raised because you are attempting to write to a socket that has been closed. This does indicate a probable bug so check your application as to why it is occurring and attempt to fix that first.
Attempting to just mask SIGPIPE is not a good idea because you don't really know where the signal is coming from and you may mask other sources of this error. In multi-threaded environments, signals are a horrible solution.
In the rare cases were you cannot avoid this, you can mask the signal on send. If you set the MSG_NOSIGNAL flag on send()/sendto(), it will prevent SIGPIPE being raised. If you do trigger this error, send() returns -1 and errno will be set to EPIPE. Clean and easy. See man send for details.

unix network process

I was wondering how tcp/ip communication is implemented in unix. When you do a send over the socket, does the tcp/level work (assembling packets, crc, etc) get executed in the same execution context as the calling code?
Or, what seems more likely, a message is sent to some other daemon process responsible for tcp communication? This process then takes the message and performs the requested work of copying memory buffers and assembling packets etc.? So, the calling code resumes execution right away and tcp work is done in parallel? Is this correct?
Details would be appreciated. Thanks!
The TCP/IP stack is part of your kernel. What happens is that you call a helper method which prepares a "kernel trap". This is a special kind of exception which puts the CPU into a mode with more privileges ("kernel mode"). Inside of the trap, the kernel examines the parameters of the exception. One of them is the number of the function to call.
When the function is called, it copies the data into a kernel buffer and prepares everything for the data to be processed. Then it returns from the trap, the CPU restores registers and its original mode and execution of your code resumes.
Some kernel thread will pick up the copy of the data and use the network driver to send it out, do all the error handling, etc.
So, yes, after copying the necessary data, your code resumes and the actual data transfer happens in parallel.
Note that this is for TCP packets. The TCP protocol does all the error handling and handshaking for you, so you can give it all the data and it will know what to do. If there is a problem with the connection, you'll notice only after a while since the TCP protocol can handle short network outages by itself. That means you'll have "sent" some data already before you'll get an error. That means you will get the error code for the first packet only after the Nth call to send() or when you try to close the connection (the close() will hang until the receiver has acknowledged all packets).
The UDP protocol doesn't buffer. When the call returns, the packet is on it's way. But it's "fire and forget", so you only know that the driver has put it on the wire. If you want to know whether it has arrived somewhere, you must figure out a way to achieve that yourself. The usual approach is have the receiver send an ack UDP packet back (which also might get lost).
No - there is no parallel execution. It is true that the execution context when you're making a system call is not the same as your usual execution context. When you make a system call, such as for sending a packet over the network, you must switch into the kernel's context - the kernel's own memory map and stack, instead of the virtual memory you get inside your process.
But there are no daemon processes magically dispatching your call. The rest of the execution of your program has to wait for the system call to finish and return whatever values it will return. This is why you can count on return values being available right away when you return from the system call - values like the number of bytes actually read from the socket or written to a file.
I tried to find a nice explanation for how the context switch to kernel space works. Here's a nice in-depth one that even focuses on architecture-specific implementation:
http://www.ibm.com/developerworks/linux/library/l-system-calls/

Resources