A rampaging mob finally convinces me to spend some quality time with DLL development...

Last week, on Wednesday, the Unidata netCDF team spent the day with Quincey and Larry of the HDF5 team. This was great because we usually don't get to spend this much time with Quincey, and we worked out a lot of issues relating to netCDF/HDF5 interoperability.

I came away with the following action items:

• switch to WEAK file close
• enable write access for HDF5 files without creation ordering
• deferred metadata read
• show multi-dimensional atts as 1D, like Java
• ignore reference types
• try to allow attributes on user defined types
• forget about stored property lists
• throw away extra links to groups and objects (like Java does)
• work with Kent/Elena on docs for NASA/GIP
• hdf4 netCDF v2 API writes as well as reads HDF4. How should this be handled?
• John suggests not using EOS libraries but just recoding that functionality.
• HDF5 team will release tool for those in big-endian wasteland. It will rewrite the file.
• should store software version in netcdf-4 file somewhere in hidden att.
• use HDF5 function to find file type, this supports user block
• read gip article
• update netCDF wikipedia page with format compatibility info
• data models document for GIP?
  

I have been assured that this blog is write-only, so I don't have to explain any of he above, because no one is reading this! ;-)

The tasks above, when complete, with together add up to a lot more interoperability between netCDF-4 and existing HDF5 data files, allowing netCDF tools to be used on HDF5 files.

This week I am attending the HDF5 workshop in Champaign, Illinois. I am learning a lot of interesting things about HDF5, and I gave a presentation on netCDF, which is now available on the netCDF web site for those that are interested:

Hartnett, E., 2010-09: NetCDF and HDF5 - HDF5 Workshop 2010.

It's great to see the HDF5 team again!

There have been many performance improvements in the upcoming netCDF-4.1.2 release.

One improvement is a complete refactor of all netCDF-4 memory structures. Now the metadata of a netCDF file occupies the smallest possible amount of memory. I have added many more Valgrind tests, and the HDF5 team has worked hard to track down memory issues in HDF5. (Most were not really bugs, but just doing things that Valrgrid doesn't like.)

It's particularly important on high performance platforms that memory used be minimized. If you run a program with 10,000 processors, and each of them uses too much memory for the metadata, that adds up to a lot of wasted memory. And in HPC they have better uses for their memory.

The biggest improvement in performance came from a rewrite of the way that netCDF-4 reads the HDF5 file. The code has been rewritten in terms of the H5LIterate() function, and this has resulted in a huge performance gain. Here's an email from Russ quantifying this gain:

From: Russ Rew <russ-AT-unidata.ucar-DOT-edu>
Subject: timings of nc_open speedup
To: ed-AT-unidata.ucar-DOT-edu
Date: Thu, 23 Sep 2010 15:23:12 -0600
Organization: UCAR Unidata Program
Reply-to: russ-AT-unidata.ucar-DOT-edu                                                                                                                                                    

Ed,

On Jennifer Adam's file, here's the before and after timings on buddy (on the file and a separate copy, to defeat caching):

  real  0m32.60s
  user  0m0.15s
  sys   0m0.46s

  real  0m0.14s
  user  0m0.01s
  sys   0m0.02s

which is a 233x speedup.

Here's before and after for test files I created that have twice as many levels as Jennifer Adam's and much better compression:

  real  0m23.78s
  user  0m0.24s
  sys   0m0.60s

  real  0m0.05s
  user  0m0.01s
  sys   0m0.01s

which is a 475x speedup.  By using even more levels, the speedup becomes arbitrarily large, because now nc_open takes a fixed amount of time that depends on the amount of metadata, not the amount of data.

--Russ

As Russ notes, this is a speedup that can be defined as arbitrarily large, if we tailor the input file correctly. But Jennifer's file is a real one, and at18.4 giga-bytes (name: T159_1978110112.nc4) this file is a real disk-buster. Yet it has a simple metadata structure. At a > 200 times speedup is nice. We had been talking about a new file open mode which would not open the file and read the metadata, all because it was taking so long. I guess I don't have to code that up now, so that's a least a couple of weeks work saved by this fix! (Not to mention that now netCDF-4 will work much better for these really big files, which are becoming more and more common.)

Here's the ncdump -h of this lovely test file:

netcdf T159_1978110112 {
dimensions:
        lon = 320 ;
        lat = 160 ;
        lev = 11 ;
        time = 1581 ;
variables:
        double lon(lon) ;
                lon:units = "degrees_east" ;
                lon:long_name = "Longitude" ;
        double lat(lat) ;
                lat:units = "degrees_north" ;
                lat:long_name = "Latitude" ;
        double lev(lev) ;
                lev:units = "millibar" ;
                lev:long_name = "Level" ;
        double time(time) ;
                time:long_name = "Time" ;
                time:units = "minutes since 1978-11-01 12:00" ;
        float temp(time, lev, lat, lon) ;
                temp:missing_value = -9.99e+08f ;
                temp:longname = "Temperature [K]" ;
                temp:units = "K" ;
        float geop(time, lev, lat, lon) ;
                geop:missing_value = -9.99e+08f ;
                geop:longname = "Geopotential [m^2/s^2]" ;
                geop:units = "m^2/s^2" ;
        float relh(time, lev, lat, lon) ;
                relh:missing_value = -9.99e+08f ;
                relh:longname = "Relative Humidity [%]" ;
                relh:units = "%" ;
        float vor(time, lev, lat, lon) ;
                vor:missing_value = -9.99e+08f ;
                vor:longname = "Vorticity [s^-1]" ;
                vor:units = "s^-1" ;
        float div(time, lev, lat, lon) ;
                div:missing_value = -9.99e+08f ;
                div:longname = "Divergence [s^-1]" ;
                div:units = "s^-1" ;
        float uwnd(time, lev, lat, lon) ;
                uwnd:missing_value = -9.99e+08f ;
                uwnd:longname = "U-wind [m/s]" ;
                uwnd:units = "m/s" ;
        float vwnd(time, lev, lat, lon) ;
                vwnd:missing_value = -9.99e+08f ;
                vwnd:longname = "V-wind [m/s]" ;
                vwnd:units = "m/s" ;
    float sfp(time, lat, lon) ;
                sfp:missing_value = -9.99e+08f ;
                sfp:longname = "Surface Pressure [Pa]" ;
                sfp:units = "Pa" ;

// global attributes:
                :NCO = "4.0.2" ;
} 

Special thanks to Jennifer Adams, from the GrADS project. Not only did she provide this great test file, but she also built my branch distribution and tested the fix for me! Thanks Jennifer! Thanks also to Quincey of HDF5 for helping me sort out the best way to read a HDF5 file.

Now I just have to make sure that parallel I/O is working OK, and then 4.1.2 will be ready for release!

A last minute change before the 4.1 release ensures that this common case will get good performance.

There is a terrible performance hit if your chunk cache is too small to hold even one chunk, and your data are deflated.

Since the default HDF5 chunk cache size is 1 MB, this is not hard to do.

So I have added code such that, when a file is opened, if the data are compressed, and if the chunksize is greater than the default chunk cache size for that var, then the chunk cache is increased to a multiple of the chunk size.

The code looks like this:

/* Is this a deflated variable with a chunksize greater than the                                                                                               
 * current cache size? */
if (!var->contiguous && var->deflate)
{
   chunk_size_bytes = 1;
   for (d = 0; d < var->ndims; d++)
     chunk_size_bytes *= var->chunksizes[d];
   if (var->type_info->size)
     chunk_size_bytes *= var->type_info->size;
   else
     chunk_size_bytes *= sizeof(char *);
#define NC_DEFAULT_NUM_CHUNKS_IN_CACHE 10
#define NC_DEFAULT_MAX_CHUNK_CACHE 67108864
   if (chunk_size_bytes > var->chunk_cache_size)
   {
     var->chunk_cache_size = chunk_size_bytes * NC_DEFAULT_NUM_CHUNKS_IN_CACHE;
     if (var->chunk_cache_size > NC_DEFAULT_MAX_CHUNK_CACHE)
        var->chunk_cache_size = NC_DEFAULT_MAX_CHUNK_CACHE;
     if ((retval = nc4_reopen_dataset(grp, var)))
        return retval;
   }
} 

I am setting the chunk cache to 10 times the chunk size, up to 64 MB max. Reasonable? Comments are welcome.

The timing results show a clear difference. First, two runs without any per-variable caching, but the second run sets a 64MB file level chunk
cache that speeds up timing considerably. (The last number in the row is the average read time for a horizontal layer, in miro-seconds.)

bash-3.2$ ./tst_ar4_3d  pr_A1_z1_256_128_256.nc 
256     128     256     1.0             1       0           836327       850607

bash-3.2$ ./tst_ar4_3d -c 68000000 pr_A1_z1_256_128_256.nc 
256     128     256     64.8            1       0           833453       3562

Without the cache it is over 200 times slower.

Now I have turned on automatic variable caches when appropriate:

bash-3.2$ ./tst_ar4_3d  pr_A1_z1_256_128_256.nc 
256     128     256     1.0             1       0           831470       3568

In this run, although no file level cache was turned on, I got the same response time. That's because when opening the file the netCDF library noticed that this deflated var had a chunk size bigger than the default cache size, and opened a bigger cache.

All of this work is in support of the general netCDF user writing very large files, and specifically in support of the AR-5 effort.

The only downside is that, if you open up a file with many such variables, and you have very little memory on your machine, you will run out of memory.