source: XIOS/dev/branch_openmp/Note/rapport ESIWACE.tex.backup @ 1552

\documentclass[a4paper,10pt]{article}
\usepackage[utf8]{inputenc}
\usepackage{graphicx}
\usepackage{listings}
\usepackage[usenames,dvipsnames,svgnames,table]{xcolor}
\usepackage{amsmath}
\usepackage{url}

% Title Page

\title{Developing XIOS with multi-thread : to accelerate the I/O of climate models}

\author{}


\begin{document}
\maketitle

\section{Context}

The simulation models of climate systems, running on a large number of computing resources can produce an important volume of data. At this 
scale, the I/O and the post-treatment of data becomes a bottle-neck for the performance. In order to manage efficiently the data flux 
generated by the simulations, we use XIOS developed by the Institut Pierre Simon Laplace and Maison de la simulation.

XIOS, a library dedicated to intense calculates, allows us to easily and efficiently manage the parallel I/O on the storage systems. XIOS 
uses the client/server scheme in which computing resources (server) are reserved exclusively for IO in order to minimize their impact on 
the performance of the climate models (client). The clients and servers are executed in parallel and communicate asynchronously. In this 
way, the I/O peaks can be smoothed out as data fluxes are send to server constantly throughout the simulation and the time spent on data 
writing on the server side can be overlapped completely by calculates on the client side. 

\begin{figure}[ht]
\includegraphics[scale=0.4]{Charge1.png}
\includegraphics[scale=0.4]{Charge2.png}
\caption{On the left, each peak of computing power corresponds to the valley of memory bandwidth which shows that the computing resources 
are alternating between calculates and I/O. ON the right, both curves are smooth which means that the computing resources have a stable 
charge of work, either calculates or I/O.}
\end{figure}


XIOS works well with many climate simulation codes. For example, LMDZ\footnote{LMDZ is a general circulation model (or global climate model) 
developed since the 70s at the "Laboratoire de Météorologie Dynamique", which includes various variants for the Earth and other planets 
(Mars, Titan, Venus, Exoplanets). The 'Z' in LMDZ stands for "zoom" (and the 'LMD' is for  'Laboratoire de Météorologie Dynamique"). 
\url{http://lmdz.lmd.jussieu.fr}}, NENO\footnote{Nucleus for European Modeling of the Ocean alias NEMO is a 
state-of-the-art modelling framework of ocean related engines. \url{https://www.nemo-ocean.eu}}, ORCHIDEE\footnote{the land surface 
model of the IPSL (Institut Pierre Simon Laplace) Earth System Model. \url{https://orchidee.ipsl.fr}}, and DYNAMICO\footnote{The DYNAMICO 
project develops a new dynamical core for LMD-Z, the atmospheric general circulation model (GCM) part of IPSL-CM Earth System Model. 
\url{http://www.lmd.polytechnique.fr/~dubos/DYNAMICO/}} all use XIOS as the output back end. M\'et\'eoFrance and MetOffice also choose XIOS 
to manege the I/O for their models.


\section{Development of thread-friendly XIOS}

Although XIOS copes well with many models, there is one potential optimization in XIOS which needs to be investigated: making XIOS thread-friendly.

This topic comes along with the configuration of the climate models. Take LMDZ as example, it is designed with the 2-level parallelization scheme. To be more specific, LMDZ uses the domain decomposition method in which each sub-domain is associated with one MPI process. Inside of the sub-domain, the model also uses OpenMP derivatives to accelerate the computation. We can imagine that the sub-domain be divided into sub-sub-domain and is managed by threads. 

\begin{figure}[ht]
\centering
\includegraphics[scale=0.5]{domain.pdf}
\caption{Illustration of the domain decomposition used in LMDZ.}
\end{figure}

As we know, each sub-domain, or in another word, each MPI process is a XIOS client. The data exchange between client and XIOS servers is handled by MPI communications. In order to write an output field, all threads must gather the data to the master thread who acts as MPI process in order to call MPI routines. There are two disadvantages about this method : first, we have to spend time on gathering information to the master thread which not only increases the memory use, but also implies an OpenMP barrier; second, while the master thread calls MPI routine, other threads are in the idle state thus a waster of computing resources. What we want obtain with the thread-friendly XIOS is that all threads can act like MPI processes. They can call directly the MPI routine thus no waste in memory nor in computing resources as shown in Figure \ref{fig:omp}.

\begin{figure}[ht]
\centering
\includegraphics[scale=0.6]{omp.pdf}
\caption{}
\label{fig:omp}
\end{figure}

There are two ways to make XIOS thread-friendly. First of all, change the structure of XIOS which demands a lot of modification is the XIOS library. Knowing that XIOS is about 100 000 lines of code, this method will be very time consuming. What's more, the modification will be local to XIOS. If we want to optimize an other code to be thread-friendly, we have to redo the modifications. The second choice is to add an extra interface to MPI in order to manage the threads. When a thread want to call an MPI routine inside XIOS, it will first pass the interface, in which the communication information will be analyzed before the MPI routine is invoked. With this method, we only need to modify a very small part of XIOS in order to make it work. What is more interesting is that the interface we created can be adjusted to suit other MPI based libraries.


In this project, we choose to implement an interface to handle the threads. To do so, we introduce the MPI\_endpoint which is a 
concept proposed in the last MPI Forums and several papers have already discussed the importance of such idea and have introduced the 
framework of the MPI\_endpoint \cite{Dinan:2013}\cite{Sridharan:2014}. The concept of an endpoint is shown by Figure \ref{fig:scheme}. In 
the MPI\_endpoint environment, each OpenMP thread will be associated with a unique rank (global endpoint rank), an endpoint communicator, 
and a local rank (rank inside the MPI process) which is very similar to the \verb|OMP_thread_num|. The global endpoint rank will replace the 
role of the classic MPI rank and will be used in MPI communication calls. 


\begin{figure}[ht]
\begin{center}
\includegraphics[scale=0.4]{scheme.png} 
\end{center}
\caption{}
\label{fig:scheme}
\end{figure}

%XIOS is a library dedicated to IO management of climate code. It has a client-server pattern in which clients are in charge of computations and servers manage the reading and writing of files. The communication between clients and servers are handled by MPI. However, some of the climate models (\textit{e.g.} LMDZ) nowadays use an hybrid programming policy. Within a shared memory node, OpenMP directives are used to manage message exchanges. In such configuration, XIOS can not take full advantages of the computing resources to maximize the performance. This is because XIOS can only work with MPI processes. Before each call of XIOS routines, threads of one MPI process must gather their information to the master thread who works as an MPI process. After the call, the master thread distributes the updated information among its slave threads. As result, all slave threads have to wait while the master thread calls the XIOS routines. This introduce extra synchronization into the model and leads to not optimized performance. Aware of this situation, we need to develop a new version of XIOS (EP\_XIOS) which can work with threads, or in other words, can consider threads as they were processes. To do so, we introduce the MPI endpoints.  

An other important aspect about the MPI\_endpoint interface is that each endpoints has knowledge of the ranks of other endpoints in the 
same communicator. This knowledge is necessary because when executing an MPI communication, for example a point-to-point exchange, we need 
to know not only the ranks of sender/receiver threads, but also the thread number of the sender/receiver threads and the MPI ranks 
of the sender/receiver processes. This ranking information is implemented inside an map object included in the endpoint communicator 
class.




%\newpage

%The MPI\_endpoint interface we implemented lies on top of an existing MPI Implementation. It consists of wrappers to all MPI functions 
%used in XIOS. 


In XIOS, we used the ``probe'' technique to search for arrived messages and then perform the receiving action. The principle is 
that sender process executes the send operation as usual. However, to minimize the time spent on waiting incoming messages, the receiver 
process calls in the first place the \verb|MPI_Probe| function to check if a message destinate to it has been published. If yes, the 
process execute in the second place the \verb|MPI_Recv| to receive the message. If not, the receiver process can carry on with other tasks 
or repeats the \verb|MPI_Probe| and \verb|MPI_Recv| actions if the required message is in immediate need. This technique works well in 
the current version of XIOS. However, if we introduce threads into this mechanism, problems can occur: The incoming message is labeled by 
the tag and receiver's MPI rank. Because threads within a process share the MPI rank, and the message probed is always available in the 
message queue, it can lead to the problem of data race and thus the message can be received by the wrong thread.


To solve this problem, we introduce the ``matching-probe'' technique. The idea of the method is that each thread is equipped with a local 
incoming message queue. Each time a thread calls an MPI function, for example \verb|MPI_Recv|, it calls firstly the \verb|MPI_Mprobe| 
function to query the MPI incoming message with any tag and from any source. Once a message is probed, the thread gets the handle to the 
incoming message and this specific message is erased from the MPI message queue. Then, the thread proceed the identification of the message 
to get the destination thread's local rank and store the message handle to the local queue of the target thread. The thread repeats these 
steps until the MPI incoming message queue is empty. Then the thread we perform the usual ``probe'' technique to query its local incoming 
message queue to check if the required message is available. If yes, it performs the \verb|MPI_Recv| operation. With this ``matching-probe'' 
technique, we can assure that a message is probed only once and is received by the right receiver.
 


Another issue needs to be clarified with this technique is that: how to identify the receiver's rank? The solution to this question is to 
use the tag argument. In the MPI environment, a tag is an integer ranging from 0 to $2^{31}$ depending on the Implementation. We can explore 
the large range property of the tag to store in it information about the source and destination thread ranks. In our endpoint interface, we 
choose to limit the first 15 bits for the tag used in the classic MPI communication, the next 8 bits to store the sender thread's local 
rank, and the last 8 bits to store the receiver thread's local rank (\textit{c.f.} Figure \ref{fig:tag}). In this way, with an extra 
analysis of the tag, we can identify the local ranks of the sender and the receiver in any P2P communication. As results, when a thread 
probes a message, it knows exactly in which local queue should store the probed message. 

\begin{figure}[ht]
 \centering
 \includegraphics[scale=0.4]{tag.png}
 \caption{}\label{fig:tag}
\end{figure}

In Figure \ref{fig:tag}, Tag contains the user defined value for a certain communication. MPI\_tag is computed in the endpoint interface 
with help of the rank map and is used in the MPI calls. 

\begin{figure}[ht]
\centering
 \includegraphics[scale = 0.4]{sendrecv.png}
\caption{This figure shows the classic pattern of a P2P communication with the endpoint interface. Thread/endpoint rank 0 sends a message 
to thread/endpoint rank 3 with tag=1. The underlying MPI function called by the sender is indeed a send for MPI rank of 1 
and tag=65537. From the receiver's point of view, the endpoint 3 is actually receving a message from MPI rank 0 with 
tag=65537.}
\label{fig:sendrecv}
\end{figure}




With the rank map, tag extension, and the matching-probe techniques, we are now able to call any P2P communication in the endpoint 
environment. For the collective communications, we apply a step-by-step execution pattern and no special technique is required. A 
step-by-step execution pattern consists of 3 steps (not necessarily in this order and not all steps are needed): arrangement of the source 
data, execution of the MPI function by all master/root threads, distribution or arrangement of the resulting data among threads. 

%The most representative functions of the collective communications are \verb|MPI_Gather| and \verb|MPI_Bcast|.

For example, if we want to perform a broadcast operation, only 2 steps are needed (\textit{c.f.} Figure \ref{fig:bcast}). Firstly, the root 
thread, along with the master threads of other processes, perform the classic \verb|MPI_Bcast| operation. Secondly, the root thread, and the 
master threads send data to other threads via local memory transfer. 

\begin{figure}[ht]
\centering
\includegraphics[scale=0.3]{bcast.png} 
\caption{}
\label{fig:bcast}
\end{figure}

Figure \ref{fig:allreduce} illustrates how the \verb|MPI_Allreduce| function is proceeded in the endpoint interface. First of all, We 
perform a intra-process ``allreduce'' operation: source data is reduced from slave threads to the master thread via local memory transfer. 
Next, alm master threads call the classic \verb|MPI_Allreduce| routine. Finally, all master threads send the updated reduced data to its 
slaves via local memory transfer. 

\begin{figure}[ht]
\centering
\includegraphics[scale=0.3]{allreduce.png} 
\caption{}
\label{fig:allreduce}
\end{figure}

Other MPI routines, such as \verb|MPI_Wait|, \verb|MPI_Intercomm_create| \textit{etc.}, can be found in the technique report of the 
endpoint interface.

\section{The multi-threaded XIOS and performce results}

The development of endpoint interface for thread-friendly XIOS library took about one year and a half. The main difficulty is the 
co-existance of MPI processes and OpenMP threads. All MPI classes must be redefined in the endpoint interface along with all the routines. 
The development is now available on the forge server: \url{http://forge.ipsl.jussieu.fr/ioserver/browser/XIOS/dev/branch_openmp}. One 
technique report is also available in which one can find more detail about how endpoint works and how the routines are implemented 
\cite{ep:2018}. We must note that the thread-friendly XIOS library is still in the phase of optimization. It will be released in the 
future with a stable version.

All the funcionalities of XIOS is reserved in its thread-friendly version. Single threaded code can work successfully with the new 
version of XIOS. For multi-threaded models, some modifications are needed in order to work with the multi-threaded XIOS library. Detail can 
be found in our technique report \cite{ep:2018}.

Even though the multi-threaded

\section{Performance of LMDZ using EP\_XIOS}

With the new version of XIOS, we are now capable of taking full advantages of the computing resources allocated by a simulation model when 
calling XIOS functions. All threads, can participate in XIOS as if they are MPI processes. We have tested the EP\_XIOS in LMDZ and the 
performance results are very encouraging.

In our tests, we used 12 client processor with 8 threads each (96 XIOS clients in total), and one single-thread server processor. We have 2 
output densities. The light output gives mainly 2 dimensional fields while the heavy output records more 3D fields. We also have differente 
simulation duration settings: 1 day, 5 days, 15 days, and 31 days. 

\begin{figure}[ht]
 \centering
 \includegraphics[scale = 0.6]{LMDZ_perf.png}
 \caption{Speedup obtained by using EP in LMDZ simulations.}
\end{figure}

In this figure, we show the speedup which is computed by $\displaystyle{\frac{time_{XIOS}}{time_{EP\_XIOS}}}$. The blue bars 
represent speedup of the XIOS file output and the red bars the speedup of LMDZ: calculates + XIOS file output. In all experimens, 
we can observe a speedup which represents a gain in performance. One important conclusion we can get from this result is that, more dense 
the output is, more efficient is the EP\_XIOS. With 8 threads per process, we can reach a speedup in XIOS upto 6, and a speedup of 1.5 in 
LMDZ which represents a decrease of the total execution time to 68\% ($\approx 1/1.5$). This observation confirmes steadily the importance 
of using EP in XIOS.  

The reason why LMDZ does not show much speedup, is because the model is calcutation dominant: time spent on calculation is much longer than 
that on the file output. For example, if 30\% of the execution time is spent on the output, then with a speepup of 6, we can obtain a 
decrease in time of 25\%. Even the 25\% may seems to be small, it is still a gain in performance with existing computing resources.

\section{Performance of EP\_XIOS}

workfloz\_cmip6 
light output
24*8+2
30s - 52s
32 days
histmth with daily output

\section{Perspectives of EP\_XIOS}


\bibliographystyle{plain}
\bibliography{reference}

\end{document}