Return to the "Advanced Topics" page.



Metadata Caching in HDF5

1. Introduction

In the 1.6.4 release, we introduced a re-implementation of the metadata cache. That release contained an incomplete version of the cache which could not be controlled via the API. The version in the 1.8 release is more mature, and includes new API calls that allow the user program to configure the metadata cache both on file open and at run time.

From the user perspective, the most striking effect of the new cache should be a large reduction in the cache memory requirements when working with complex HDF5 files.

Those working with such files may also notice a reduction in file close time.

Those working with HDF5 files with simple structure shouldn't notice any particular changes in most cases. In rare cases, there may be a significant improvement in performance.

The remainder of this document contains an architectural overview of the old and new metadata caches, a discussion of algorithms used to automatically adjust cache size to circumstances, and a high level discussion of the cache configuration controls. It can be safely skipped by anyone who works only with HDF5 files with relatively simple structure (i.e. no huge groups, no datasets with large numbers of chunks, and no objects with large numbers of attributes.)

On the other hand, it is mandatory reading if you want to use something other than the default metadata cache configuration. The documentation on the metadata cache related API calls will not make much sense without this background.

2. Old and New Metadata Cache

2.1. The Old Metadata Cache

The old metadata cache indexed the cache with a hash table with no provision for collisions. Instead, collisions were handled by evicting the existing entry to make room for the new entry. Aside from flushes, there was no other mechanism for evicting entries, so the replacement policy could best be described as "Evict on Collision".

As a result, if two frequently used entries hashed to the same location, they would evict each other regularly. To decrease the likelihood of this situation, the default hash table size was set fairly large -- slightly more than 10,000. This worked well, but since the size of metadata entries is not bounded, and since entries were only evicted on collision, the large hash table size allowed the cache size to explode when working with HDF5 files with complex structure.

The "Evict on Collision" replacement policy also caused problems with the parallel version of the HDF5 library, as a collision with a dirty entry could force a write in response to a metadata read. Since all metadata writes must be collective in the parallel case while reads need not be, this could cause the library to hang if only some of the processes participated in a metadata read that forced a write. Prior to the implementation of the new metadata cache, we dealt with this issue by maintaining a shadow cache for dirty entries evicted by a read.

2.2. The New Metadata Cache

The new metadata cache was designed to address the above issues. After implementation, it became evident that the working set size for HDF5 files varies widely depending on both structure and access pattern. Thus it was necessary to add support for cache size adjustment under either automatic or user program control (see section 2.3 for details).

When the cache is operating under direct user program control, it is also possible to temporarily disable evictions from the metadata cache so as to maximize raw data throughput at the expense of allowing the cache to grow without bound until evictions are enabled again.

Structurally, the new metadata cache can be thought of as a heavily modified version of the UNIX buffer cache as described in chapter three of M. J. Bach's "The Design of the UNIX Operating System" In essence, the UNIX buffer cache uses a hash table with chaining to index a pool of fixed size buffers. It uses the LRU replacement policy to select candidates for eviction.

Since HDF5 metadata entries are not of fixed size, and may grow arbitrarily large, the size of the new metadata cache cannot be controlled by setting a maximum number of entries. Instead the new cache keeps a running sum of the sizes of all entries, and attempts to evict entries as necessary to stay within a user specified maximum size. (Note the use of the word "attempts" here -- as will be seen, it is possible for the cache to exceed its currently specified maximum size.) At present, the LRU replacement policy is the only option for selecting candidates for eviction.

Per the standard unix buffer cache, dirty entries are given two passes through the LRU list before being evicted. The first time they reach the end of the LRU list, they are flushed, marked as clean, and moved to the head of the LRU list. When a clean entry reaches the end of the LRU list, it is simply evicted if space is needed.

The cache cannot evict entries that are locked, and thus it will temporarily grow beyond its maximum size if there are insufficient unlocked entries available for eviction.

In the parallel version of the library, only the cache running under process 0 of the file communicator is allowed to write metadata to file. All the other caches must retain dirty metadata until the process 0 cache tells them that the metadata is clean.

Since all operations modifying metadata must be collective, all caches see the same stream of dirty metadata. This fact is used to allow them to synchronize every n bytes of dirty metadata, where n is a user configurable value that defaults to 256 KB.

To avoid sending the other caches messages from the future, process 0 must not write any dirty entries until it reaches a synchronization point. When it reaches a synchronization point, it writes entries as needed, and then broadcasts the list of flushed entries to the other caches. The caches on the other processes use this list to mark entries clean before they leave the synchronization point, allowing them to evict those entries as needed.

The caches will also synchronize on a user initiated flush.

To minimize overhead when running in parallel, the cache maintains a "clean" LRU list in addition to the regular LRU list. This list contains only clean entries, and is used as a source of candidates for eviction when flushing dirty entries is not allowed.

Since flushing entries is forbidden most of the time when running in parallel, the caches can be forced to exceed their maximum sizes if they run out of clean entries to evict.

To decrease the likelihood of this event, the new cache allows the user to specify a minimum clean size -- which is a minimum total size of all the entries on the clean LRU plus all unused space in the cache.

While the clean LRU list is only maintained in the parallel version of the HDF5 library, the notion of a minimum clean size still applies in the serial case. Here it is used to force a mix of clean and dirty entries in the cache even in the write only case.

This in turn reduces the number of redundant flushes by avoiding the case in which the cache fills with dirty metadata and all entries must be flushed before a clean entry can be evicted to make room for a new entry.

Observe that in both the serial and parallel cases, the maintenance of a minimum clean size modifies the replacement policy, as dirty entries may be flushed earlier than would otherwise be the case so as to maintain the desired amount of clean and/or empty space in the cache.

While the new metadata cache only supports the LRU replacement policy at present, that may change. Support for multiple replacement policies was very much in mind when the cache was designed, as was the ability to switch replacement policies at run time. The situation has been complicated by the later addition of the adaptive cache resizing requirement, as two of the resizing algorithms piggyback on the LRU list. However, if there is need for additional replacement policies, it shouldn't be too hard to implement them.

3. Adaptive Cache Resizing in the New Metadata Cache

As mentioned earlier, the metadata working set size for a HDF5 file varies wildly depending on the structure of the file and the access pattern. For example, a 2MB limit on metadata cache size is excessive for an H5repack of almost all HDF5 files we have tested. However, I have a file submitted by one of our users that that will run a 13% hit rate with this cache size, and will lock up one of our linux boxes using the old metadata cache. Increase the new metadata cache size to 4 MB, and the hit rate exceeds 99%.

In this case the main culprit is a root group with more than 20,000 entries in it. As a result, the root group heap exceeds 1 MB, which tends to crowd out the rest of the metadata in a 2 MB cache

This case and a number of synthetic tests convinced us that we needed to modify the new metadata cache to expand and contract according to need within user specified bounds.

I was unable to find any previous work on this problem, so I invented solutions as I went along. If you are aware of prior work, please send me references. The closest I was able to come was a group of embedded CPU designers who were turning off sections of their cache to conserve power.

3.1. Increasing the Cache Size

In the context of the HDF5 library, the problem of increasing the cache size as necessary to contain the current working set turns out to involve two rather different issues.

The first of these, which was recognized immediately, is the problem of recognizing long term changes in working set size, and increasing the cache size accordingly, while not reacting to transients.

The second, which I recognized the hard way, is to adjust the cache size for sudden, dramatic increases in working set size caused by requests for large pieces of metadata which may be larger than the current metadata cache size.

The algorithms for handling these situations are discussed below. These problems are largely orthogonal to each other, so both algorithms may be used simultaneously.

3.1.1. Hit Rate Threshold Cache Size Increment

Perhaps the most obvious heuristic for identifying cases in which the cache is too small involves monitoring the hit rate. If the hit rate is low for a while, and the cache is at its current maximum size, the current maximum cache size is probably too small.

The hit rate threshold algorithm for increasing cache size applies this intuition directly.

Hit rate statistics are collected over a user specified number of cache accesses. This period is known as an epoch.

At the end of each epoch, the hit rate is computed, and the counters are reset. If the hit rate is below a user specified threshold and the cache is at its current maximum size, the maximum size of the cache is increased by a user specified multiple. If required, the new cache maximum size is clipped to stay within the user specified upper bound on the maximum cache size, and optionally, within a user specified maximum increment.

My tests indicate that this algorithm works well in most cases. However, in a synthetic test in which hit rate increased slowly with cache size, and load remained steady for many epochs, I observed a case in which cache size increased until hit rate just exceeded the specified minimum and then stalled. This is a problem, as to avoid volatility, it is necessary to set the minimum hit rate threshold well below the desired hit rate. Thus we may find ourselves with a cache running with a 91% hit rate when we really want it to increase its size until the hit rate is about 99%.

If this case occurs frequently in actual use, I will have to come up with an improved algorithm. Please let me know if you see this behavior. However, I had to work rather hard to create it in my synthetic tests, so I would expect it to be uncommon.

3.1.2. Flash Cache Size Increment

A fundamental problem with the above algorithm is that contains the hidden assumption that cache entries are relatively small in comparison to the cache itself. While I knew this assumption was not generally true when I developed the algorithm, I thought that cases where it failed would be so rare as to not be worth considering, as even if they did occur, the above algorithm would rectify the situation within an epoch or two.

While it is true that such occurances are rare, and it is true that the hit rate threshold cache size increment algorithm will rectify the situation eventually, the performance degradation experienced by users while waiting for the epoch to end was so extreme that some way of accelerating response to such situations was essential.

To understand the problem, consider the following use case:

Suppose we create a group, and then repeatedly create a new data set in the group, write some data to it and then close it.

In some versions of the HDF5 file format, the names of the datasets will be stored in a local heap associated with the group, and the space for that heap will be allocated in a single, contiguous chunk. When this local heap is full, we allocate a new chunk twice the size of the old, copy the data from the old local heap into the new, and discard the old local heap.

By default, the minimum metadata cache size is set to 2 MB. Thus in this use case, our hit rate will be fine as long as the local heap is no larger than a little less than 2 MB, as the group related metadata is accessed frequently and never evicted, and the data set related metadata is never accessed once the data set is closed, and thus is evicted smoothly to make room for new data sets.

All this changes abruptly when the local heap finally doubles in size to a value above the slightly less than 2 MB limit. All of a sudden, the local heap is the size of the metadata cache, and the cache must constantly swap it in to access it, and then swap it out to make room for other metadata.

The hit rate threshold based algorithm for increasing the cache size will fix this problem eventually, but performance will be very bad until it does, as the metadata cache will largely ineffective until its size is increase.

An obvious heuristic for addressing this "big rock in a small pond" issue is to watch for large "incoming rocks", and increase the size of the "pond" if the rock is so big that it will force most of the "water" out of the "pond".

The add space flash cache size increment algorithm algorithm applies this intuition directly:

Let x be either the size of a newly inserted entry, a newly loaded entry, or the number of bytes by which the size of an existing entry has been increased (i.e. the size of the "rock").

If x is greater than some user specified fraction of the current maximum cache size, increase the current maximum cache size by x times some user specified multiple, less any free space that was in the cache to begin with. Further, to avoid confusing the other cache size increment/decrement code, start a new epoch.

At present, this algorithm pays no attention to any user specified limit on the maximum size of any single cache size increase, but it DOES stay within the user specified upper bound on the maximum cache size.

While it should be easy to see how this algorithm could be fooled into inactivity by large number of entries that were not quite large enough to cross the threshold, in practice it seems to work reasonably well.

Needless to say, I will revisit the issue should this cease to be the case.

3.2. Decreasing the Cache Size

Identifying cases in which the maximum cache size is larger than necessary turned out to be more difficult.

3.2.1. Hit Rate Threshold Cache Size Reduction

One obvious heuristic is to monitor the hit rate and guess that we can safely decrease cache size if hit rate exceeds some user supplied threshold (say .99995).

The hit rate threshold size decrement algorithm implemented in the new metadata cache implements this intuition as follows:

At the end of each epoch (this is the same epoch that is used in the cache size increment algorithm), the hit rate is compared with the user specified threshold. If the hit rate exceeds that threshold, the current maximum cache size is decreased by a user specified factor. If required, the size of the reduction is clipped to stay within a user specified lower bound on the maximum cache size, and optionally, within a user specified maximum decrement.

In my synthetic tests, this algorithm works poorly. Even with a very high threshold and a small maximum reduction, it results in cache size oscillations. The size increment code typically increments maximum cache size above the working set size. This results in a high hit rate, which causes the threshold size decrement code to reduce the maximum cache size below the working set size, which causes hit rate to crash causing the cycle to repeat. The resulting average hit rate is poor.

It remains to be seen if this behavior will be seen in the field. The algorithm is available for use, but it wouldn't be my first choice. If you use it, please report back.

3.2.2. Ageout Cache Size Reduction

Another heuristic for dealing with oversized cache conditions is to look for entries that haven't been accessed for a long time, evict them, and reduce the cache size accordingly.

The age out cache size reduction applies this intuition as follows: At the end of each epoch (again the same epoch as used in the cache size increment algorithm), all entries that haven't been accessed for a user configurable number of epochs (1 - 10 at present) are evicted. The maximum cache size is then reduced to equal the sum of the sizes of the remaining entries. The size of the reduction is clipped to stay within a user specified lower bound on maximum cache size, and optionally, within a user specified maximum decrement.

In addition, the user may specify a minimum fraction of the cache which must be empty before the cache size is reduced. Thus if an empty reserve of 0.1 was specified on a 10 MB cache, there would be no cache size reduction unless the eviction of aged out entries resulted in more than 1 MB of empty space. Further, even after the reduction, the cache would be one tenth empty.

In my synthetic tests, the age out algorithm works rather well, although it is somewhat sensitive to the epoch length and age out period selection.

3.2.3. Ageout With Hit Rate Threshold Cache Size Reduction

To address these issues, I combined the hit rate threshold and age out heuristics.

Age out with threshold works just like age out, except that the algorithm is not run unless the hit rate exceeded a user specified threshold in the previous epoch.

In my synthetic tests, age out with threshold seems to work nicely, with no observed oscillation. Thus I have selected it as the default cache size reduction algorithm.

For those interested in such things, the age out algorithm is implemented by inserting a marker entry at the head of the LRU list at the beginning of each epoch. Entries that haven't been accessed for at least n epochs are simply entries that appear in the LRU list after the n-th marker at the end of an epoch.

4. Configuring the New Metadata Cache

Due to lack of resources, the design work on the automatic cache size adjustment algorithms was done hastily, using primarily synthetic tests. I don't think I spent more than a couple weeks writing and running performance tests -- most time went into coding and functional testing.

As a result, while I think the algorithms provided for adaptive cache resizing will work well in actual use, I don't really know (although preliminary results from the field are promising). Fortunately, the issue shouldn't arise for the vast majority of HDF5 users, and those for whom it may arise should be savvy enough to recognize problems and deal with them.

For this latter class of users, I have implemented a number of new API calls allowing the user to select and configure the cache resize algorithms, or to turn them off and control cache size directly from the user program. There are also API calls that allow the user program to monitor hit rate and cache size.

From the user perspective, all the cache configuration data for a given file is contained in an instance of the H5AC_cache_config_t structure -- the definition of which is given below:

        typedef struct H5AC_cache_config_t
        {
            /* general configuration fields: */
            int                         version;

            hbool_t                     rpt_fcn_enabled;

            hbool_t                     open_trace_file;
            hbool_t                     close_trace_file;
            char                        trace_file_name
                                          [H5AC__MAX_TRACE_FILE_NAME_LEN + 1];

            hbool_t                     evictions_enabled;

            hbool_t                     set_initial_size;
            size_t                      initial_size;

            double                      min_clean_fraction;

            size_t                      max_size;
            size_t                      min_size;

            long int                    epoch_length;


            /* size increase control fields: */
            enum H5C_cache_incr_mode    incr_mode;

            double                      lower_hr_threshold;

            double                      increment;

            hbool_t                     apply_max_increment;
            size_t                      max_increment;

            enum H5C_cache_flash_incr_mode flash_incr_mode;
            double                      flash_multiple;
	    double                      flash_threshold;


            /* size decrease control fields: */
            enum H5C_cache_decr_mode    decr_mode;

            double                      upper_hr_threshold;

            double                      decrement;

            hbool_t                     apply_max_decrement;
            size_t                      max_decrement;

            int                         epochs_before_eviction;

            hbool_t                     apply_empty_reserve;
            double                      empty_reserve;


            /* parallel configuration fields: */
            int                         dirty_bytes_threshold;

        } H5AC_cache_config_t;

This structure is defined in H5ACpublic.h. Each field is discussed below and in the associated header comment.

The C API allows you get and set this structure directly. Unfortunately the Fortran API has to do this with individual parameters for each of the fields (with the exception of version).

While the API calls are discussed individually in the reference manual, the following high level discussion of what fields to change for different purposes should be useful.

4.1. General Configuration

The version field is intended to allow THG to change the H5AC_cache_config_t structure without breaking old code. For now, this field should always be set to H5AC__CURR_CACHE_CONFIG_VERSION, even when you are getting the current configuration data from the cache. The library needs the version number to know where fields are located with reference to the supplied base address.

The rpt_fcn_enabled field is a boolean flag that allows you to turn on and off the resize reporting function that reports the activities of the adaptive cache resize code at the end of each epoch -- assuming that it is enabled.

The report function is unsupported, so you are on your own if you use it. Since it dumps status data to stdout, you should not attempt to use it with Windows unless you modify the source. You may find it useful if you want to experiment with different adaptive resize configurations. It is also a convenient way of diagnosing poor cache configuration. Finally, if you do lots of runs with identical behavior, you can use it to determine the metadata cache size needed in each phase of your program so you can set the required cache sizes manually.

The trace file fields are also unsupported. They allow one to open and close a trace file in which all calls to the metadata cache are logged in a user specified file for later analysis. The feature is intended primarily for THG use in debugging or optimizing the metadata cache in cases where users in the field observe obscure failures or poor performance that we cannot re-create in the lab. The trace file will allow us to re-create the exact sequence of cache operations that are triggering the problem.

At present we do not have a play back utility for trace files, although I imagine that we will write one quickly when and if we need it.

To enable the trace file, you load the full path of the desired trace file into trace_file_name, and set open_trace_file to TRUE. In the parallel case, an ASCII representation of the rank of each process is appended to the supplied trace file name to create a unique trace file name for that process.

To close an open trace file, set close_trace_file to TRUE.

It must be emphasized that you are on your own if you play with the trace file feature absent a request from THG. Needless to say, the trace file feature is disabled by default. If you enable it, you will take a large performance hit and generate huge trace files.

The evictions_enabled field is a boolean flag allowing the user to disable the eviction of entries from the metadata cache. Under normal operation conditions, this field will always be set to TRUE.

In rare circumstances, the raw data throughput requirements may be so high that the user wishes to postpone metadata writes so as to reserve I/O throughput for raw data. The evictions_enabled field exists to allow this -- although the user is to be warned that the metadata cache will grow without bound while evictions are disabled. Thus evictions should be re-enabled as soon as possible, and it may be wise to monitor cache size and statistics (to see how to enable statistics, see the debugging facilities section below).

Evictions may only be disabled when the automatic cache resize code is disabled as well. Thus to disable evictions, not only must the user set the evictions_enabled field to FALSE, but he must also set incr_mode to H5C_incr__off, set flash_incr_mode to H5C_flash_incr__off, and set decr_mode to H5C_decr__off.

To re-enable evictions, just set evictions_enabled back to TRUE.

Before passing on to other subjects, it is worth re-iterating that disabling evictions is an extreme step. Before attempting it, you might consider setting a large cache size manually, and flushing the cache just before high raw data throughput is required. This may yield the desired results without the risks inherent in disabling evictions.

The set_initial_size and initial_size fields allow you to specify an initial maximum cache size. If set_initial_size is TRUE, initial_size must lie in the interval [min_size, max_size] (see below for a discussion of the min_size and max_size fields).

If you disable the adaptive cache resizing code (done by setting incr_mode to H5C_incr__off, flash_incr_mode to H5C_flash_incr__off, and decr_mode to H5C_decr__off), you can use these fields to control maximum cache size manually, as the maximum cache size will remain at the initial size.

Note, that the maximum cache size is only modified when set_initial_size is TRUE. This allows the use of configurations specified at compile time to change resize configuration without altering the current maximum size of the cache. Without this feature, an additional call would be required to get the current maximum cache size so as to set the initial_size to the current maximum cache size, and thereby avoid changing it.

The min_clean_fraction sets the current minimum clean size as a fraction of the current max cache size. While this field was originally used only in the parallel version of the library, it now applies to the serial version as well. Its value must lie in the range [0.0, 1.0]. 0.01 is reasonable in the serial case, and 0.3 in the parallel.

A potential interaction, discovered at release 1.8.3, between the enforcement of the min_clean_fraction and the adaptive cache resize code can severely degrade performance. While this interaction is easily dealt in the serial case by setting min_clean_fraction to 0.01, the problem is more difficult in the parallel case. Please see the “Interactions” section below for further details.

The max_size and min_size fields specify the range of maximum sizes that may be set for the cache by the automatic resize code. min_size must be less than or equal to max_size, and both must lie in the range [H5C__MIN_MAX_CACHE_SIZE, H5C__MAX_MAX_CACHE_SIZE] -- currently [1 KB, 128 MB]. If you routinely run a cache size in the top half of this range, you should increase the hash table size. To do this, modify the H5C__HASH_TABLE_LEN #define in H5Cpkg.h and re-compile. At present, H5C__HASH_TABLE_LEN must be a power of two.

The epoch_length is the number of cache accesses between runs of the adaptive cache size control algorithms. It is ignored if these algorithms are turned off. It must lie in the range [H5C__MIN_AR_EPOCH_LENGTH, H5C__MAX_AR_EPOCH_LENGTH] -- currently [100, 1000000]. The above constants are defined in H5Cprivate.h. 50000 is a reasonable value.

4.2. Increment Configuration

The incr_mode field specifies the cache size increment algorithm used. Its value must be a member of the H5C_cache_incr_mode enum type -- currently either H5C_incr__off or H5C_incr__threshold (note the double underscores after "incr"). This type is defined in H5Cpublic.h.

If incr_mode is set to H5C_incr__off, regular automatic cache size increases are disabled, and the lower_hr_threshold, increment, apply_max_increment, and max_increment fields are ignored.

The flash_incr_mode field specifies the flash cache size increment algorithm used. Its value must be a member of the H5C_cache_flash_incr_mode enum type -- currently either H5C_flash_incr__off or H5C_flash_incr__add_space (note the double underscores after "incr"). This type is defined in H5Cpublic.h.

If flash_incr_mode is set to H5C_flash_incr__off, flash cache size increases are disabled, and the flash_multiple, and flash_threshold, fields are ignored.

4.2.1. Hit Rate Threshold Cache Size Increase Configuration

If incr_mode is H5C_incr__threshold, the cache size is increased via the hit rate threshold algorithm. The remaining fields in the section are then used as follows:

lower_hr_threshold is the threshold below which the hit rate must fall to trigger an increase. The value must lie in the range [0.0 - 1.0]. In my tests, a relatively high value seems to work best -- 0.9 for example.

increment is the factor by which the old maximum cache size is multiplied to obtain an initial new maximum cache size when an increment is needed. The actual change in size may be smaller as required by max_size (above) and max_increment (discussed below). increment must be greater than or equal to 1.0. If you set it to 1.0, you will effectively turn off the increment code. 2.0 is a reasonable value.

apply_max_increment and max_increment allow the user to specify a maximum increment. If apply_max_increment is TRUE, the cache size will never be increased by more than the number of bytes specified in max_increment in any single increase.

4.2.2. Flash Cache Size Increase Configuration

If flash_incr_mode is set to H5C_flash_incr__add_space, flash cache size increases are enabled. The size of the cache will be increased under the following circumstances:

Let t be the current maximum cache size times the value of the flash_threshold field.

Let x be either the size of the newly inserted entry, the size of the newly loaded entry, or the number of bytes added to the size of the entry under consideration for triggering a flash cache size increase.

If t < x, the basic condition for a flash cache size increase is met, and we proceed as follows:

Let space_needed equal x less the amount of free space in the cache.

Further, let increment equal space_needed times the value of the flash_multiple field. If increment plus the current cache size is greater than max_size (discussed above), reduce increment so that increment plus the current cache size is equal to max_size.

If increment is greater than zero, increase the current cache size by increment. To avoid confusing the other cache size increment or decrement algorithms, start a new epoch. Note however, that we do not cycle the epoch markers if some variant of the age out algorithm is in use.

The use of the flash_threshold field is discussed above. It must be a floating point value in the range of [0.1, 1.0]. 0.25 is a reasonable value.

The use of the flash_multiple field is also discussed above. It must be a floating point value in the range of [0.1, 10.0]. 1.4 is a reasonable value.

4.3. Decrement Configuration

The decr_mode field specifies the cache size decrement algorithm used. Its value must be a member of the H5C_cache_decr_mode enum type -- currently either H5C_decr__off, H5C_decr__threshold, H5C_decr__age_out, or H5C_decr__age_out_with_threshold (note the double underscores after "decr"). This type is defined in H5Cpublic.h.

If decr_mode is set to H5C_decr__off, automatic cache size decreases are disabled, and the remaining fields in the cache size decrease control section are ignored.

4.3.1. Hit Rate Threshold Cache Size Decrease Configuration

if decr_mode is H5C_decr__threshold, the cache size is decreased by the threshold algorithm, and the remaining fields of the decrement section are used as follows:

upper_hr_threshold is the threshold above which the hit rate must rise to trigger cache size reduction. It must be in the range [0.0, 1.0]. In my synthetic tests, very high values like .9995 or .99995 seemed to work best.

decrement is the factor by which the current maximum cache size is multiplied to obtain a tentative new maximum cache size. It must lie in the range [0.0, 1.0]. Relatively large values like .9 seem to work best in my synthetic tests. Note that the actual size reduction may be smaller as required by min_size and max_decrement (discussed below).

apply_max_decrement and max_decrement allow the user to specify a maximum decrement. If apply_max_decrement is TRUE, cache size will never be reduced by more than max_decrement bytes in any single reduction.

With the hit rate threshold cache size decrement algorithm, the remaining fields in the section are ignored.

4.3.2. Ageout Cache Size Reduction

If decr_mode is H5C_decr__age_out the cache size is decreased by the ageout algorithm, and the remaining fields of the decrement section are used as follows:

epochs_before_eviction is the number of epochs an entry must reside unaccessed in the cache before it is evicted. This value must lie in the range [1, H5C__MAX_EPOCH_MARKERS]. H5C__MAX_EPOCH_MARKERS is defined in H5Cprivate.h, and is currently set to 10.

apply_max_decrement and max_decrement are used as in section 2.4.3.1.

apply_empty_reserve and empty_reserve allow the user to specify a minimum empty reserve as discussed in section 2.3.2.2. An empty reserve of 0.05 or 0.1 seems to work well.

The decrement and upper_hr_threshold fields are ignored in this case.

4.3.3. Ageout With Hit Rate Threshold Cache Size Reduction

If decr_mode is H5C_decr__age_out_with_threshold, the cache size is decreased by the ageout with hit rate threshold algorithm, and the fields of decrement section are used as per the Ageout algorithm (see 5.3.2) with the exception of upper_hr_threshold.

Here, upper_hr_threshold is the threshold above which the hit rate must rise to trigger cache size reduction. It must be in the range [0.0, 1.0]. In my synthetic tests, high values like .999 seemed to work well.

4.4. Parallel Configuration

This section is a catch-all for parallel specific configuration data. At present, it has only one field -- dirty_bytes_threshold.

In PHDF5, all operations that modify metadata must be executed collectively. We used to think that this was enough to ensure consistency across the metadata caches, but since we allow processes to read metadata individually, the order of dirty entries in the LRU list can vary across processes. This in turn can change the order in which dirty metadata cache entries reach the bottom of the LRU and are flushed to disk -- opening the door to messages from the past and messages from the future bugs.

To prevent this, only the metadata cache on process 0 of the file communicator is allowed to write to file, and then only after entering a sync point with the other caches. After it writes entries to file, it sends the base addresses of the now clean entries to the other caches, so they can mark these entries clean as well, and then leaves the sync point. The other caches mark the specified entries as clean before they leave the synch point as well. (Observe, that since all caches see the same stream of dirty metadata, they will all have the same set of dirty entries upon sync point entry and exit.)

The different caches know when to synchronize by counting the number of bytes of dirty metadata created by the collective operations modifying metadata. Whenever this count exceeds the value specified in the dirty_bytes_threshold, they all enter the sync point, and process 0 flushes down to its minimum clean size and sends the list of newly cleaned entries to the other caches.

Needless to say, the value of the dirty_bytes_threshold field must be consistent across all the caches operating on a given file.

All dirty metadata can also by flushed under programatic control via the H5Fflush() call. This call must be collective, and will reset the dirty data counts on each metadata cache.

Absent calls to H5Fflush(), dirty metadata will only be flushed when the dirty_bytes_threshold is exceeded, and then only down to the min_clean_fraction. Thus, if a program does all its metadata modifications in one phase, and then doesn't modify metadata thereafter, a residue of dirty metadata will be frozen in the metadata caches for the remainder of the computation -- effectively reducing the sizes of the caches.

In the default configuration, the caches will eventually resize themselves to maintain an acceptable hit rate. However, this will take time, and it will increase the applications footprint in memory.

If your application behaves in this manner, you can avoid this by a collective call to H5Fflush() immediately after the metadata modification phase.

4.5. Interactions

Evictions may not be disabled unless the automatic cache resize code is disabled as well (by setting decr_mode to H5C_decr__off, flash_decr_mode to H5C_flash_incr__add_space, and incr_mode to H5C_incr__off) -- thus placing the cache size under the direct control of the user program.

There is no logical necessity for this restriction. It is imposed because it simplifies testing greatly, and because I can't see any reason why one would want to disable evictions while the automatic cache size adjustment code was enabled. This restriction can be relaxed if anyone can come up with a good reason to do so.

At present there are two interactions between the increment and decrement sections of the configuration.

If incr_mode is H5C_incr__threshold, and decr_mode is either H5C_decr__threshold or H5C_decr__age_out_with_threshold, then lower_hr_threshold must be strictly less than upper_hr_threshold.

Also, if the flash cache size increment code is enabled and is triggered, it will restart the current epoch without calling any other cache size increment or decrement code.

In both the serial and parallel cases, there is the potential for an interaction between the min_clean_fraction and the cache size increment code that can severly degrade performance. Specifically, if the min_clean_fraction is large enough, it is possible that keeping the specified fraction of the cache clean may generate enough flushes to seriously degrade performance even though the hit rate is excellent.

In the serial case, this is easily dealt with by selecting a very small min_clean_fraction -- 0.01 for example -- as this still avoids the "metadata blizzard" phenomenon that appears when the cache fills with dirty metadata and must then flush all of it before evicting an entry to make space for a new entry.

The problem is more difficult in the parallel case, as the min_clean_fraction is used ensure that the cache contains clean entries that can be evicted to make space for new entries when metadata writes are forbidden -- i.e. between sync points.

This issue was discovered shortly before release 1.8.3 and an automated solution has not been implemented. Should it become an issue for an application, try manually setting the cache size to ~1.5 times the maximum working set size for the application, and leave min_clean_fraction set to 0.3.

You can approximate the working set size of your application via repeated calls to H5Fget_mdc_size() and H5Fget_mdc_hit_rate() while running your program with the cache resize code enabled. The maximum value returned by H5Fget_mdc_size() should be a reasonable approximation -- particularly if the associated hit rate is good.

In the parallel case, there is also an interaction between min_clean_fraction and dirty_bytes_threshold. Absent calls to H5Fflush() (discussed above), the upper bound on the amount of dirty data in the metadata caches will oscillate between (1 - min_clean_fraction) times current maximum cache size, and that value plus the dirty_bytes_threshold. Needless to say, it will be best if the min_size, min_clean_fraction, and the dirty_bytes_threshold are chosen so that the cache can't fill with dirty data.

4.6. Default Metadata Cache Configuration

Starting with release 1.8.3, HDF5 provides different default metadata cache configurations depending on whether the library is compiled for serial or parallel.

The default configuration for the serial case is as follows:

{
  /* int         version                = */ H5C__CURR_AUTO_SIZE_CTL_VER,
  /* hbool_t     rpt_fcn_enabled        = */ FALSE,
  /* hbool_t     open_trace_file        = */ FALSE,
  /* hbool_t     close_trace_file       = */ FALSE,
  /* char        trace_file_name[]      = */ "",
  /* hbool_t     evictions_enabled      = */ TRUE,
  /* hbool_t     set_initial_size       = */ TRUE,
  /* size_t      initial_size           = */ ( 2 * 1024 * 1024),
  /* double      min_clean_fraction     = */ 0.01,
  /* size_t      max_size               = */ (32 * 1024 * 1024),
  /* size_t      min_size               = */ ( 1 * 1024 * 1024),
  /* long int    epoch_length           = */ 50000,
  /* enum H5C_cache_incr_mode incr_mode = */ H5C_incr__threshold,
  /* double      lower_hr_threshold     = */ 0.9,
  /* double      increment              = */ 2.0,
  /* hbool_t     apply_max_increment    = */ TRUE,
  /* size_t      max_increment          = */ (4 * 1024 * 1024),
  /* enum H5C_cache_flash_incr_mode       */
  /*                    flash_incr_mode = */ H5C_flash_incr__add_space,
  /* double      flash_multiple         = */ 1.4,
  /* double      flash_threshold        = */ 0.25,
  /* enum H5C_cache_decr_mode decr_mode = */ H5C_decr__age_out_with_threshold,
  /* double      upper_hr_threshold     = */ 0.999,
  /* double      decrement              = */ 0.9,
  /* hbool_t     apply_max_decrement    = */ TRUE,
  /* size_t      max_decrement          = */ (1 * 1024 * 1024),
  /* int         epochs_before_eviction = */ 3,
  /* hbool_t     apply_empty_reserve    = */ TRUE,
  /* double      empty_reserve          = */ 0.1,
  /* int         dirty_bytes_threshold  = */ (256 * 1024)
}

The default configuration for the parallel case is as follows:

{
  /* int         version                = */ H5C__CURR_AUTO_SIZE_CTL_VER,
  /* hbool_t     rpt_fcn_enabled        = */ FALSE,
  /* hbool_t     open_trace_file        = */ FALSE,
  /* hbool_t     close_trace_file       = */ FALSE,
  /* char        trace_file_name[]      = */ "",
  /* hbool_t     evictions_enabled      = */ TRUE,
  /* hbool_t     set_initial_size       = */ TRUE,
  /* size_t      initial_size           = */ ( 2 * 1024 * 1024),
  /* double      min_clean_fraction     = */ 0.3,
  /* size_t      max_size               = */ (32 * 1024 * 1024),
  /* size_t      min_size               = */ ( 1 * 1024 * 1024),
  /* long int    epoch_length           = */ 50000,
  /* enum H5C_cache_incr_mode incr_mode = */ H5C_incr__threshold,
  /* double      lower_hr_threshold     = */ 0.9,
  /* double      increment              = */ 2.0,
  /* hbool_t     apply_max_increment    = */ TRUE,
  /* size_t      max_increment          = */ (4 * 1024 * 1024),
  /* enum H5C_cache_flash_incr_mode       */
  /*                    flash_incr_mode = */ H5C_flash_incr__add_space,
  /* double      flash_multiple         = */ 1.0,
  /* double      flash_threshold        = */ 0.25,
  /* enum H5C_cache_decr_mode decr_mode = */ H5C_decr__age_out_with_threshold,
  /* double      upper_hr_threshold     = */ 0.999,
  /* double      decrement              = */ 0.9,
  /* hbool_t     apply_max_decrement    = */ TRUE,
  /* size_t      max_decrement          = */ (1 * 1024 * 1024),
  /* int         epochs_before_eviction = */ 3,
  /* hbool_t     apply_empty_reserve    = */ TRUE,
  /* double      empty_reserve          = */ 0.1,
  /* int         dirty_bytes_threshold  = */ (256 * 1024)
}

The default serial configuration should be adequate for most serial HDF5 users.

The same may not be true for the default parallel configuration due the interaction between the min_clean_fraction and the cache size increase code. See the “Interactions” section for further details.

Should you need to change the default configuration, it can be found in H5ACprivate.h. Look for the definition of H5AC__DEFAULT_RESIZE_CONFIG.

5. Controlling the New Metadata Cache Size From Your Program

You have already seen how H5AC_cache_config_t has facilities that allow you to control the metadata cache size directly. Use H5Fget_mdc_config() and H5Fset_mdc_config() to get and set the metadata cache configuration on an open file. Use H5Pget_mdc_config() and H5Pset_mdc_config() to get and set the initial metadata cache configuration in a file access property list. Recall that this list contains configuration data used when opening a file.

Use H5Fget_mdc_hit_rate() to get the average hit rate since the last time the hit rate stats were reset. This happens automatically at the beginning of each epoch if the adaptive cache resize code is enabled. You can also do it manually with H5Freset_mdc_hit_rate_stats(). Be careful about doing this if the adaptive cache resize code is enabled, as you may confuse it.

Use H5Fget_mdc_size() to get metadata cache size data on an open file.

Finally, note that cache size and cache footprint are two different things -- in my tests, the cache footprint (as inferred from the UNIX top command) is typically about three times the maximum cache size. I haven't tracked it down yet, but I would guess that most of this is due to the very small typical cache entry size combined with the rather large size of cache entry header structure. This should be investigated further, but there are other matters of higher priority.

6. New Metadata Cache Debugging Facilities

The new metadata cache has a variety of debugging facilities that may be of use. I doubt that any other than the report function and the trace file will ever be accessible via the API, but they are relatively easy to turn on in the source code.

Note that none of this should be viewed as supported -- it is described here on the off chance that you want to use it, but you are on your own if you do. Also, there are no promises as to consistency between versions.

As mentioned above, you can use the rpt_fcn_enabled field of the configuration structure to enable the default reporting function (H5C_def_auto_resize_rpt_fcn() in H5C.c). If this function doesn't work for you, you will have to write your own. In particular, remember that it uses stdout, so it will probably be unhappy under Windows.

Again, remember that this facility is not supported. Further, it is likely to change every time I do any serious work on the cache.

There is also extensive statistics collection code. Use H5C_COLLECT_CACHE_STATS and H5C_COLLECT_CACHE_ENTRY_STATS in H5Cprivate.h to turn this on. If you also turn on H5AC_DUMP_STATS_ON_CLOSE in H5ACprivate.h, stats will be dumped when you close a file. Alternatively you can call H5C_stats() and H5C_stats__reset() within the library to dump and reset stats. Both of these functions are defined in H5C.c.

Finally, the cache also contains extensive sanity checking code. Much of this is turned on when you compile in debug mode, but to enable the full suite, turn on H5C_DO_SANITY_CHECKS in H5Cprivate.h

7. Trouble Shooting

Absent major bugs in the cache, the only trouble shooting you should have to do is diagnosing and fixing problems with your cache configuration.

Assuming it runs on your platform (I've only used it under Linux), the reporting function is probably the most convenient diagnosis tool. However, since it is unsupported code, I will not discuss it further beyond directing you to the source (H5C_def_auto_resize_rpt_fcn() in H5C.c).

Absent the reporting function, regular calls to H5Fget_mdc_hit_rate() should give you a good idea of hit rate over time. Remember that the hit rate stats are reset at the end of each epoch (when adaptive cache resizing is enabled), so you should expect some jitter.

Similar calls to H5Fget_mdc_size() should allow you to monitor cache size, and the fraction of the current maximum cache size that is actually in use.

If the hit rate is consistently low, and the cache it at its current maximum size, increasing the maximum size is an obvious fix.

If you see hit rate and cache size oscillations, try disabling adaptive cache resizing and setting a fixed cache size a bit greater than the high end of the cache size oscillations you observed.

If the hit rate oscillations don't go away, you are probably looking at a feature of your application which can't be helped without major changes to the cache. Please send along a description of the situation.

If the oscillations do go away, you may be able to come up with a configuration that deals with the situation. If that fails, control cache size manually, and write me, so I can try to develop an adaptive resize algorithm that works in your case.

Needless to say, you should give the cache a few epochs to adapt to circumstances. If that is too slow for you, try manual cache size control.

If you find it necessary to disable evictions, you may find it useful to enable the internal statistics collection code mentioned above in the section on debugging facilities.

Amongst many other other things, the stats code will report the the maximum cache size, and the average successful and unsuccessful search depths in the hash table. If these latter figures are significantly above 1, you should increase the size of the hash table.



Return to the "Advanced Topics" page.




The HDF Group Help Desk:
Describes HDF5 Release 1.8.7, targeted for May 2011.