Some tips of using “pool” in programming

In the past weeks, I am dived into utilizing CUDA APIs to operate on multiple GPUs. Among my work, one is a “memory pool” module which manages allocating/freeing memories from different devices. In this post, I want to share some thoughts about “pool”, an interesting tech in programming.

The “pool” works like a “cache”: it allocates some resource beforehand; if the application wants it, the “pool” picks up a available one for it. One classical example is the “database connection pool”: the “pool” preallocates some TCP connections and keep them alive, and this will save client from handshaking with server every time. The other instance is “memory pool”, which I implemented recently. The “pool” keeps the freed memory, and not really release it to device. Based on my benchmark test, the application’s performance can get a 100% improvement in extreme case.

The other function of using “pool” is for debugging purpose. Still use my “memory pool” as an demonstration: for every memory pointer, there are a device ID (which GPU this memory is allocated from) and memory size accompanied with it. So you can know the whole life of this block memory clearly from analyzing trace log. This has saved me from notorious “an illegal memory access was encountered” error many times in multiple GPU programming.

Last but not least, although the number of this “memory pool”‘s code lines is merely ~100, I already use following data structures: queue, vector, pair and map. Not mention the mutex and lock, which are essential to avoid nasty data-race issues. So writing this model is a good practice to hone my programming craft.

Enjoy coding!

An empirical method of debugging “illegal memory access” bug in CUDA programming

Recently, I encountered “an illegal memory access was encountered” error during CUDA programming. such as:

cudaSafeCall(cudaMemcpy(h_res, d_res, gParams.n*sizeof(uint32), cudaMemcpyDeviceToHost));

Because the kernel in CUDA programming is executed asynchronously, the code which reports error is not the original culprit usually. After referring this piece of code, I encapsulate a new cudaMemoryTest function:

#define cudaSafeCall(call)  \
        do {\
            cudaError_t err = call;\
            if (cudaSuccess != err) \
            {\
                std::cerr << "CUDA error in " << __FILE__ << "(" << __LINE__ << "): " \
                    << cudaGetErrorString(err);\
                exit(EXIT_FAILURE);\
            }\
        } while(0)

void cudaMemoryTest()
{
    const unsigned int N = 1048576;
    const unsigned int bytes = N * sizeof(int);
    int *h_a = (int*)malloc(bytes);
    int *d_a;
    cudaSafeCall(cudaMalloc((int**)&d_a, bytes));

    memset(h_a, 0, bytes);
    cudaSafeCall(cudaMemcpy(d_a, h_a, bytes, cudaMemcpyHostToDevice));
    cudaSafeCall(cudaMemcpy(h_a, d_a, bytes, cudaMemcpyDeviceToHost));

    return 0;
}

And insert this function call to the suspicious positions:

{
    cudaMemoryTest();
    kernel_A<<<...>>>;

    ......

    cudaMemoryTest();
    kernel_B<<<...>>>;

    ......
}

This method can notify me timely once the CUDA memory access is exceptional, then I can investigate further.

Mostly, the reasons causing this issue is NULL pointer or a pointer points to a already freed memory. But in multiple GPUs environment, you must make sure the memory for one operation is allocated in the same device.

The time spent on practising white board test may not be worthy

Once upon a time, I spent my spare time in practising white board test, i.e., try to solve data structure / algorithm issues from leetcodehackerrank, etc. But after nearly one or two months, I decided to stop this activity since I find the time spent on it is not worthy, at least for me.

I am not a clever software engineer, even a little dumb. Except very basic problems, I find a lot of puzzles which has “medium” difficulty level will cause me spend hours to fix it. I mean although I can resolve part of them, but I am not sure whether this is the optimal solution, so I would read others’ answers which claims consume very little time carrefully to understand them. For some extreame cases which I could’t get a clue after two hours, it usually costs me to take much longer time to get the point of the puzzles.

Gradually, I found though this practice can have some effect in improving my coding ability, whereas compared to the time consumed, it is not cost-effective, and honestly not very helpful for my daily work. Take C++ as an example, the STL and boost have provided almost all the data structures and algorithms which you need. You just need to call std::sort, and don’t care it is “bubble sort” or “quick sort” under the hood. So in reality, I seldom have the opportunity to implement the sort algorithm from scratch. Furthermore, most problems just need single-thread to resolve, but I find multi-thread and synchronization are real needs in daily life. Unfortunately, concurreny programming ability can’t be polished through resloving these puzzles. Last but not least, researching these problems really consume too much time. If I really need the realted knowledge of one specified problem at work, studying it then is not too late.

After stopping these practice, I leverage the spare time for following tasks: read C++ classical books and learn more about standard libraries (There is another example, hash table is a commonly used data structure, and STL provides a lot of member functions for it, but I find many puzzles related to hash table only need very few methods, so you can’t get a comprehensive understanding of hash table class in STL by resolving these problems only); watch techncical conference videos (this can expand horizon and improve my English skills) and study basic knowledges for computet and maths (Definitely, the data structures and algorithms are important, I think we should know the internals well, but no need implementing them ourselvels).

To wrap up, currently I feel the time is fully utilized and more worthy compared to practise white board tests before. What’s your opinions? Feedback welcome!

Function application in Haskell

As a newbie of Haskell, I find the life becomes easier once I understand function application:

(1) function application is actually “function call”. For example, define a simple add function who returns the sum of 2 numbers:

# cat add.hs
add :: Num a => a -> a -> a
add a b = a + b

Load it in ghci, and call this function:

# ghci
GHCi, version 8.2.2: http://www.haskell.org/ghc/  :? for help
Prelude> :l add
[1 of 1] Compiling Main             ( add.hs, interpreted )
Ok, one module loaded.
*Main> add 2 4
6
*Main> add 3 6
9

Beware that the tokens in function application are separated by space. So once you see following format:

a b ..

You know it is a function application, and also a “function call”.

(2) function application has the highest precedence. Check following example:

*Main> add 1 2 ^ add 1 2
27

It is equal to “(add 1 2) ^ (add 1 2)” literally.

(3) $ operator is “application operator”, and it is right associative, and has lowest precedence. Check following instance:

*Main> add 1 $ add 2 $ add 3 4
10

The $ operator divides the expression into 3 parts: “add 1“, “add 2” and add 3 4. Because $ is right associative, the result of add 3 4 is fed into add 2function first; then the result of add 2 $ add 3 4 is passed into add 1. It is equal to “add 1 ( add 2 ( add 3 4 ) )” in fact, so $ can be used to remove parentheses.

References:
Prelude;
Calling functions.

Porting software is fun and rewarding

Regarding to port software, I think there are several kinds:

a) For the simplest case, one tool is created for Linux, and you want to use it on FreeBSD. Because there is no out-of-box package for this Operating System, you grab the code and compile it yourself, no complaint from compiler. Run it and it seems work, bingo! This should be a perfect experience!

b) The life will become pleasant if everything is similar to the above case, but in reality it is definitely not. Sometimes, the progress can’t go so smoothly. Take socket programming as an example, the Solaris has some specific requirements if you are only familiar with Linux environment (Please check this post). So you may tweak the compiler options and even customoize your code to fit your requirement in this scenario.

c) The third case is you need to read the whole software source code and do modifications, and this is what I am currently doing. Back to this Monday, I received a task to verify a conception. I remembered there is an Open Source framework which has implemented similar function, so I downloaded and went through the code carefully. Fortunately, this project indeed satisfies our requirement, but since our computation environment is Nvidia GPU, I need to use CUDA APIs to replace the related code besides integrate this framework into our code repository. If no other accidents, I think I can finish the whole work in next week.

From my personal experience, porting software is really rewarding! Take this week’s work as an example, I learnt a new C++ library and refreshed my knowledge of graph data structure. Furthermore, porting software can also give you fun: after several hours even days’ hard work, a bespoken tool can meet your requirement, that will let you feel very filled!

At the end, I must declare I don’t encourage you should be lazy and don’t think problems yourself; instead you should leverage the resource rationally. Moreover, please conform to the software license, and don’t violate it.

Enjoy porting!