Skip to content

Commit 3822871

Browse files
wkliaowkliao
committed
doc: add more comparison of pnetcdf and netcdf4
1 parent 0b2b144 commit 3822871

File tree

1 file changed

+336
-1
lines changed

1 file changed

+336
-1
lines changed

‎doc/netcdf4_vs_pnetcdf.md‎

Lines changed: 336 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,46 @@
1-
## Programming difference between NetCDF4 and PnetCDF
1+
# Comparison of PnetCDF and NetCDF4
22

3+
* [Supported File Formats](#supported-file-formats)
4+
* [Programming Differences](#programming-differences)
5+
* [Define Mode and Data Mode](#define-mode-and-data-mode)
6+
* [Collective and Independent I/O Mode](#collective-and-independent-io-mode)
7+
* [Blocking vs. Nonblocking APIs](#blocking-vs-nonblocking-apis)
8+
9+
---
10+
11+
## Supported File Formats
12+
* NetCDF4 supports both classic and HDF5-based file formats.
13+
+ Classic file format (CDF-1) -- The ESDS Community Standard defined the file format
14+
to be used in the NetCDF user community in 1989. The file header bears a
15+
signature of character string 'CDF-1' and now is commonly referred to as
16+
[CDF-1](https://parallel-netcdf.github.io/doc/c-reference/pnetcdf-c/CDF_002d1-file-format-specification.html)
17+
file format.
18+
* 'CDF-2' format -- The CDF-1 format was later extended to support large
19+
file size (i.e. larger than 2GB) in 2004. See its specification in
20+
[ESDS-RFC-011v2.0](https://cdn.earthdata.nasa.gov/conduit/upload/496/ESDS-RFC-011v2.00.pdf).
21+
Because its file header bears a signature of 'CDF-2' and the format is
22+
also commonly referred to as
23+
[CDF-2](https://parallel-netcdf.github.io/doc/c-reference/pnetcdf-c/CDF_002d2-file-format-specification.html)
24+
format.
25+
* [CDF-5](https://parallel-netcdf.github.io/doc/c-reference/pnetcdf-c/CDF_002d5-file-format-specification.html)
26+
format -- The CDF-2 format was extended by PnetCDF developer team
27+
in 2009 to support large variables and additional large data types, such
28+
as 64-bit integer.
29+
+ HDF5-based file format -- Starting from its version 4.0.0, NetCDF includes
30+
the format that is based on HDF5, which is referred to as NetCDF-4 format.
31+
This offer new features such as groups, compound types, variable length
32+
arrays, new unsigned integer types, etc.
33+
* PnetCDF supports only the classic file formats.
34+
+ The classic files created by applications using NetCDF4 library can be read
35+
by the PnetCDF library and vice versa.
36+
+ PnetCDF provides parallel I/O for accessing files in the classic format.
37+
NetCDF4's parallel I/O for classic files makes use of PnetCDF library
38+
underneath. Such feature can be enabled when building NetCDF4 library.
39+
40+
41+
---
42+
43+
## Programming Differences
344
* The API names are different between NetCDF4 and PnetCDF.
445
+ For C programming, NetCDF4 uses prefix `nc_` while PnetCDF uses `ncmpi_`.
546
+ For Fortran 77 programming, NetCDF4 uses prefix `nf_` while PnetCDF uses `nfmpi_`.
@@ -35,3 +76,297 @@
3576
| ${\textsf{\color{green}nc\\_put\\_vara\\_float}}$(ncid, varid, start, count, buf); | ${\textsf{\color{blue}ncmpi\\_put\\_vara\\_float\\_all}}$(ncid, varid, start, count, buf); |
3677
| /* close file */ | |
3778
| ${\textsf{\color{green}nc\\_close}}$(ncid); | ${\textsf{\color{blue}ncmpi\\_close}}$(ncid); |
79+
80+
81+
---
82+
83+
## Define Mode and Data Mode
84+
85+
In PnetCDF, an opened file is in either define mode or data mode. Switching
86+
between the modes is done by explicitly calling `"ncmpi_enddef()"` and
87+
`"ncmpi_redef()"`. NetCDF4 when operating on an HDF5-based file has no such
88+
mode switching requirement. The reason of PnetCDF enforcing such a requirement
89+
is to ensure the metadata consistency across all the MPI processes and keep the
90+
overhead of metadata synchronization small.
91+
92+
* Define mode
93+
+ When calling `"ncmpi_create()"` to create a new file, the file is
94+
automatically put in the define mode. While in the define mode, the user
95+
program can create new dimensions, new variables, and netCDF attributes.
96+
Modification of these data objects' metadata can only be done when the file
97+
is in the define mode.
98+
+ When opening an existing file, the opened file is automatically put in the
99+
data mode. To add or modify the metadata, the user program must call
100+
`"ncmpi_redef()"`.
101+
102+
* Data mode
103+
+ Once the creation or modification of metadata is complete, the user program
104+
must call `"ncmpi_enddef()"` to leave the define mode and enter the data
105+
mode.
106+
+ While an open file is in data mode, the user program can make read and
107+
write requests to that variables that have been created.
108+
109+
<ul>
110+
<li> A PnetCDF example codes below show switching between define and data
111+
modes after creating a new file.</li>
112+
<li> <details>
113+
<summary>Example code fragment (click to expand)</summary>
114+
115+
```c
116+
#include <mpi.h>
117+
#include <pnetcdf.h>
118+
...
119+
/* Create the file */
120+
ncmpi_create(MPI_COMM_WORLD, filename, NC_CLOBBER, MPI_INFO_NULL, &ncid);
121+
122+
...
123+
/* Define dimensions */
124+
ncmpi_def_dim(ncid, "Y", 16, &dimid[0]);
125+
ncmpi_def_dim(ncid, "X", 32, &dimid[1]);
126+
127+
/* Define a 2D variable of integer type */
128+
ncmpi_def_var(ncid, "grid", NC_INT, 2, dimid, &varid);
129+
130+
/* Add an attribute of string type to the variable */
131+
str_att = "example attribute of type text";
132+
ncmpi_put_att_text(ncid, varid, "str_att_name", strlen(str_att), str_att);
133+
134+
/* Exit the define mode */
135+
ncmpi_enddef(ncid);
136+
137+
/* Write to a subarray of the variable */
138+
MPI_Offset start[2], count[2];
139+
start[0] = 4;
140+
start[1] = 8;
141+
count[0] = 10;
142+
count[1] = 10;
143+
ncmpi_put_vara_int_all(ncid, varid, start, count, buf_int);
144+
145+
/* Re-enter the define mode */
146+
ncmpi_redef(ncid);
147+
148+
/* Define a new 2D variable of float type */
149+
ncmpi_def_var(ncid, "temperature", NC_FLOAT, 2, dimid, &var_flt);
150+
151+
/* Exit the define mode */
152+
ncmpi_enddef(ncid);
153+
154+
/* Write to a subarray of the variable, var_flt */
155+
start[0] = 2;
156+
start[1] = 8;
157+
count[0] = 5;
158+
count[1] = 5;
159+
ncmpi_put_vara_float_all(ncid, var_flt, start, count, buf_flt);
160+
161+
/* Close the file */
162+
ncmpi_close(ncid);
163+
```
164+
</details></li>
165+
166+
<li> An example shows switching between define and data modes after opening an existing file.
167+
</li>
168+
<li> <details>
169+
<summary>Example code fragment (click to expand)</summary>
170+
171+
```c
172+
#include <mpi.h>
173+
#include <pnetcdf.h>
174+
...
175+
/* Opening an existing file */
176+
ncmpi_open(MPI_COMM_WORLD, filename, NC_NOWRITE, MPI_INFO_NULL, &ncid);
177+
178+
...
179+
/* get the ID of variable named 'grid', a 2D variable of integer type */
180+
ncmpi_inq_varid(ncid, "grid", &varid);
181+
182+
/* Read the variable's attribute named "str_att_name" */
183+
char str_att[64];
184+
ncmpi_get_att_text(ncid, varid, "str_att_name", str_att);
185+
186+
/* Read a subarray of the variable, var */
187+
MPI_Offset start[2], count[2];
188+
start[0] = 4;
189+
start[1] = 8;
190+
count[0] = 10;
191+
count[1] = 10;
192+
ncmpi_get_vara_int_all(ncid, varid, start, count, buf_int);
193+
194+
/* Re-enter the define mode */
195+
ncmpi_redef(ncid);
196+
197+
/* Define a new 2D variable of double type */
198+
ncmpi_def_var(ncid, "precipitation", NC_DOUBLE, 2, dimid, &var_dbl);
199+
200+
/* Add an attribute of string type to the variable */
201+
str_att = "mm/s";
202+
ncmpi_put_att_text(ncid, var_dbl, "unit", strlen(str_att), str_att);
203+
204+
/* Exit the define mode */
205+
ncmpi_enddef(ncid);
206+
207+
/* Write to a subarray of the variable, var_dbl */
208+
start[0] = 2;
209+
start[1] = 8;
210+
count[0] = 5;
211+
count[1] = 5;
212+
ncmpi_put_vara_double_all(ncid, var_dbl, start, count, buf_dbl);
213+
214+
/* Close the file */
215+
ncmpi_close(ncid);
216+
```
217+
</details></li>
218+
</ul>
219+
220+
221+
---
222+
## Collective and Independent I/O Mode
223+
224+
The terminology of collective and independent I/O comes from MPI standard. A
225+
collective I/O function call requires all the MPI processes opening the same
226+
file to participate. On the other hand, an independent I/O function can be
227+
called by an MPI process independently from others.
228+
229+
For metadata I/O, both PnetCDF and NetCDF4 require the function calls to be
230+
collective.
231+
232+
* Mode Switch Mechanism
233+
+ PnetCDF -- when a file is in the data mode, it can be put into either
234+
collective or independent I/O mode. The default mode is collective I/O
235+
mode. Switching to and exiting from the independent I/O mode is done by
236+
explicitly calling `"ncmpi_begin_indep_data()"` and
237+
`"ncmpi_end_indep_data()"`.
238+
239+
+ NetCDF4 -- collective and independent mode switching is done per variable
240+
basis. Switching mode is done by explicitly calling `"nc_var_par_access()"`
241+
before accessing the variable. For more information, see
242+
[Parallel I/O with NetCDF-4](https://docs.unidata.ucar.edu/netcdf-c/current/parallel_io.html).
243+
244+
<ul>
245+
<li> A PnetCDF example shows switching between collective and independent I/O
246+
modes.</li>
247+
<li> <details>
248+
<summary>Example code fragment (click to expand)</summary>
249+
250+
```c
251+
#include <mpi.h>
252+
#include <pnetcdf.h>
253+
...
254+
/* Create the file */
255+
ncmpi_create(MPI_COMM_WORLD, filename, NC_CLOBBER, MPI_INFO_NULL, &ncid);
256+
257+
...
258+
/* Metadata operations to define dimensions and variables */
259+
...
260+
/* Exit the define mode (by default, into the collective I/O mode) */
261+
ncmpi_enddef(ncid);
262+
263+
/* Write to variables collectively */
264+
ncmpi_put_vara_int_all(ncid, varid, start, count, buf_int);
265+
266+
/* Read from variables collectively */
267+
ncmpi_get_vara_float_all(ncid, var_flt, start, count, buf_flt);
268+
269+
/* Leaving collective I/O mode and entering independent I/O mode */
270+
ncmpi_begin_indep_data(ncid);
271+
272+
/* Write to variables independently */
273+
ncmpi_put_vara_int(ncid, varid, start, count, buf_int);
274+
275+
/* Read from variables independently */
276+
ncmpi_get_vara_float(ncid, var_flt, start, count, buf_flt);
277+
278+
/* Close the file */
279+
ncmpi_close(ncid);
280+
```
281+
</details></li>
282+
</ul>
283+
284+
<ul>
285+
<li> A NetCDF4 example shows switching between collective and
286+
independent I/O modes.</li>
287+
<li> <details>
288+
<summary>Example code fragment (click to expand)</summary>
289+
290+
```c
291+
#include <mpi.h>
292+
#include <netcdf.h>
293+
#include <netcdf_par.h>
294+
...
295+
/* Create the file */
296+
nc_create_par(filename, NC_CLOBBER, MPI_COMM_WORLD, MPI_INFO_NULL, &ncid);
297+
298+
...
299+
/* Metadata operations to define dimensions and variables */
300+
...
301+
302+
/* set the access method to use MPI collective I/O for all variables */
303+
nc_var_par_access(ncid, NC_GLOBAL, NC_COLLECTIVE);
304+
305+
/* Write to variables collectively */
306+
nc_put_vara_int(ncid, varid, start, count, buf_int);
307+
308+
/* Read from variables collectively */
309+
nc_get_vara_float(ncid, var_flt, start, count, buf_flt);
310+
311+
/* set the access method to use MPI independent I/O for all variables */
312+
nc_var_par_access(ncid, NC_GLOBAL, NC_INDEPENDENT);
313+
314+
/* Write to variables independently */
315+
nc_put_vara_int(ncid, varid, start, count, buf_int);
316+
317+
/* Read from variables independently */
318+
nc_get_vara_float(ncid, var_flt, start, count, buf_flt);
319+
320+
/* Close the file */
321+
nc_close(ncid);
322+
```
323+
</details></li>
324+
</ul>
325+
326+
---
327+
328+
## Blocking vs Nonblocking APIs
329+
* Blocking APIs -- All NetCDF4 APIs are blocking APIs. A blocking API means the
330+
call to the API will not return until the operation is completed. For
331+
example, a call to `nc_put_var_float()` will return only when the write data
332+
has been stored at the system space, e.g. file systems. Similarly, a call to
333+
`nc_get_var_float()` will only return when the user read buffer containing
334+
the data retrieved from the file. Therefore, when a series of `put/get`
335+
blocking APIs are called, these calls will be committed by the NetCDF4
336+
library one at a time, following the same order of the calls.
337+
* Nonblocking APIs -- In addition to blocking APIs, PnetCDF provides the
338+
nonblocking version of the APIs. A nonblocking API means the call to the API
339+
will return as soon as the `put/get` request has been registered in the
340+
PnetCDF library. The commitment of the request may happen later, when a call
341+
to `ncmpi_wait_all/ncmpi_wait` is made. The nonblocking APIs are listed below.
342+
+ `ncmpi_iput_var_xxx()` - posts a nonblocking request to write to a variable.
343+
+ `ncmpi_iget_var_xxx()` - posts a nonblocking request to from from a variable.
344+
+ `ncmpi_bput_var_xxx()` - posts a nonblocking, buffered request to write to a variable.
345+
+ `ncmpi_iput_varn_xxx()` - posts a nonblocking request to write multiple subarrays to a variable.
346+
+ `ncmpi_iget_varn_xxx()` - posts a nonblocking request to read multiple subarrays from a variable.
347+
+ `ncmpi_bput_varn_xxx()` - posts a nonblocking, buffered request to write multiple subarrays to a variable.
348+
+ `ncmpi_wait_all()` - waits for nonblocking requests to complete, using collective MPI-IO.
349+
+ `ncmpi_wait()` - waits for nonblocking requests to complete, using independent MPI-IO.
350+
+ `ncmpi_attach_buff()` - Let PnetCDF to allocate an internal buffer to cache bput write requests.
351+
+ `File.detach_buff()` - Free the attached buffer.
352+
* The advantage of using nonblocking APIs is when there are many small
353+
`put/get` requests and each of them has a small amount. PnetCDF tries to
354+
aggregate and coalesce multiple registered nonblocking requests into a large
355+
one, because I/O usually performs better when the request amounts are large
356+
and contiguous. See example programs
357+
[nonblocking_write.c](../examples/C/nonblocking_write.c) and
358+
[bput_varn_int64.c](../examples/C/bput_varn_int64.c).
359+
* Table below shows the difference in C programming between using blocking
360+
and nonblocking APIs.
361+
362+
| PnetCDF Blocking APIs | PnetCDF Nonblocking APIs |
363+
|:-------|:--------|
364+
| ...<br>/* define 3 variables of NC_FLOAT type */ ||
365+
| ncmpi_def_var(ncid, "PSFC", NC_FLOAT, 2, dimid, &psfc);<br>ncmpi_def_var(ncid, "PRCP", NC_FLOAT, 2, dimid, &prcp);<br>ncmpi_def_var(ncid, "SNOW", NC_FLOAT, 2, dimid, &snow); | ditto |
366+
| ... ||
367+
| /* exit define mode and enter data mode */<br>ncmpi_enddef(ncid); | ditto |
368+
| ...<br>/* Call blocking APIs to write 3 variables to the file */ | <br>/* Call nonblocking APIs to post 3 write requests */ |
369+
| ncmpi_put_vara_float_all(ncid, psfc, start, count, buf_psfc);<br>ncmpi_put_vara_float_all(ncid, prcp, start, count, buf_prcp);<br> ncmpi_put_vara_float_all(ncid, snow, start, count, buf_snow);| ncmpi_iput_vara_float(ncid, psfc, start, count, buf_psfc, &req[0]);<br>ncmpi_iput_vara_float(ncid, prcp, start, count, buf_prcp, &req[1]);<br>ncmpi_iput_vara_float(ncid, snow, start, count, buf_snow, &req[2]);|
370+
| | /* Wait for nonblocking requests to complete */<br>ncmpi_wait_all(3, reqs, errs)|
371+
372+

0 commit comments

Comments
 (0)