Metadata Cache Configuration

Table of Contents

• General configuration fields
• Cache size increase control fields
• Cache size decrease control fields
• Parallel Configuration Fields

General configuration fields

version

Integer field containing the version number of this version of the H5AC_cache_config_t structure. Any instance of H5AC_cache_config_t passed to the cache must have a known version number, or an error will be flagged.

rpt_fcn_enabled

Boolean field used to enable and disable the default reporting function. This function is invoked every time the automatic cache resize code is run, and reports on its activities.

This is a debugging function, and should normally be turned off.

open_trace_file

Boolean field indicating whether the trace_file_name field should be used to open a trace file for the cache.

*** DEPRECATED *** Use H5Fstart/stop logging functions instead

The trace file is a debuging feature that allow the capture of top level metadata cache requests for purposes of debugging and/or optimization. This field should normally be set to FALSE, as trace file collection imposes considerable overhead.

This field should only be set to TRUE when the trace_file_name contains the full path of the desired trace file, and either there is no open trace file on the cache, or the close_trace_file field is also TRUE.

close_trace_file

Boolean field indicating whether the current trace file (if any) should be closed.

*** DEPRECATED *** Use H5Fstart/stop logging functions instead

See the above comments on the open_trace_file field. This field should be set to FALSE unless there is an open trace file on the cache that you wish to close.

trace_file_name

Full path of the trace file to be opened if the open_trace_file field is TRUE.

*** DEPRECATED *** Use H5Fstart/stop logging functions instead

In the parallel case, an ascii representation of the mpi rank of the process will be appended to the file name to yield a unique trace file name for each process.

The length of the path must not exceed H5AC__MAX_TRACE_FILE_NAME_LEN characters.

evictions_enabled

Boolean field used to either report the current evictions enabled status of the cache, or to set the cache's evictions enabled status.

In general, the metadata cache should always be allowed to evict entries. However, in some cases it is advantageous to disable evictions briefly, and thereby postpone metadata writes. However, this must be done with care, as the cache can grow quickly. If you do this, re-enable evictions as soon as possible and monitor cache size.

At present, evictions can only be disabled if automatic cache resizing is also disabled (that is, (incr_mode == H5C_incr__off ) && ( decr_mode == H5C_decr__off )). There is no logical reason why this should be so, but it simplifies implementation and testing, and I can't think of any reason why it would be desireable. If you can think of one, I'll revisit the issue. (JM)

set_initial_size

Boolean flag indicating whether the size of the initial size of the cache is to be set to the value given in the initial_size field. If set_initial_size is FALSE, the initial_size field is ignored.

initial_size

If enabled, this field contain the size the cache is to be set to upon receipt of this structure. Needless to say, initial_size must lie in the closed interval [min_size, max_size].

min_clean_fraction

double in the range 0 to 1 indicating the fraction of the cache that is to be kept clean. This field is only used in parallel mode. Typical values are 0.1 to 0.5.

max_size

Maximum size to which the cache can be adjusted. The supplied value must fall in the closed interval [MIN_MAX_CACHE_SIZE, MAX_MAX_CACHE_SIZE]. Also, max_size must be greater than or equal to min_size.

min_size

Minimum size to which the cache can be adjusted. The supplied value must fall in the closed interval [H5C__MIN_MAX_CACHE_SIZE, H5C__MAX_MAX_CACHE_SIZE]. Also, min_size must be less than or equal to max_size.

epoch_length

Number of accesses on the cache over which to collect hit rate stats before running the automatic cache resize code, if it is enabled.

At the end of an epoch, we discard prior hit rate data and start collecting afresh. The epoch_length must lie in the closed interval [H5C__MIN_AR_EPOCH_LENGTH, H5C__MAX_AR_EPOCH_LENGTH].

Cache size increase control fields

incr_mode

Instance of the H5C_cache_incr_mode enumerated type whose value indicates how we determine whether the cache size should be increased. At present there are two possible values:

• H5C_incr__off: Don't attempt to increase the size of the cache automatically.
  When this increment mode is selected, the remaining fields in the cache size increase section ar ignored.
• H5C_incr__threshold: Attempt to increase the size of the cache whenever the average hit rate over the last epoch drops below the value supplied in the lower_hr_threshold field.
  Note that this attempt will fail if the cache is already at its maximum size, or if the cache is not already using all available space.

Note that you must set decr_mode to H5C_incr__off if you disable metadata cache entry evictions.

lower_hr_threshold

Lower hit rate threshold. If the increment mode (incr_mode) is H5C_incr__threshold and the hit rate drops below the value supplied in this field in an epoch, increment the cache size by size_increment. Note that cache size may not be incremented above max_size, and that the increment may be further restricted by the max_increment field if it is enabled.

When enabled, this field must contain a value in the range [0.0, 1.0]. Depending on the incr_mode selected, it may also have to be less than upper_hr_threshold.

increment

Double containing the multiplier used to derive the new cache size from the old if a cache size increment is triggered. The increment must be greater than 1.0, and should not exceed 2.0.

The new cache size is obtained my multiplying the current max cache size by the increment, and then clamping to max_size and to stay within the max_increment as necessary.

apply_max_increment

Boolean flag indicating whether the max_increment field should be used to limit the maximum cache size increment.

max_increment

If enabled by the apply_max_increment field described above, this field contains the maximum number of bytes by which the cache size can be increased in a single re-size.

flash_incr_mode

Instance of the H5C_cache_flash_incr_mode enumerated type whose value indicates whether and by which algorithm we should make flash increases in the size of the cache to accommodate insertion of large entries and large increases in the size of a single entry.

The addition of the flash increment mode was occasioned by performance problems that appear when a local heap is increased to a size in excess of the current cache size. While the existing re-size code dealt with this eventually, performance was very bad for the remainder of the epoch.

At present, there are two possible values for the flash_incr_mode:

• H5C_flash_incr__off: Don't perform flash increases in the size of the cache.

• H5C_flash_incr__add_space: Let x be either the size of a newly newly inserted entry, or the number of bytes by which the size of an existing entry has been increased.
  If x > flash_threshold * current max cache size, increase the current maximum cache size by x * flash_multiple less any free space in the cache, and star a new epoch. For now at least, pay no attention to the maximum increment.

In both of the above cases, the flash increment pays no attention to the maximum increment (at least in this first incarnation), but DOES stay within max_size.

With a little thought, it should be obvious that the above flash cache size increase algorithm is not sufficient for all circumstances – for example, suppose the user round robins through (1/flash_threshold) +1 groups, adding one data set to each on each pass. Then all will increase in size at about the same time, requiring the max cache size to at least double to maintain acceptable performance, however the above flash increment algorithm will not be triggered.

Hopefully, the add space algorithms detailed above will be sufficient for the performance problems encountered to date. However, we should expect to revisit the issue.

flash_multiple

Double containing the multiple described above in the H5C_flash_incr__add_space section of the discussion of the flash_incr_mode section. This field is ignored unless flash_incr_mode is H5C_flash_incr__add_space.

flash_threshold

Double containing the factor by which current max cache size is multiplied to obtain the size threshold for the add_space flash increment algorithm. The field is ignored unless flash_incr_mode is H5C_flash_incr__add_space.

Cache size decrease control fields

decr_mode

Instance of the H5C_cache_decr_mode enumerated type whose value indicates how we determine whether the cache size should be decreased. At present there are four possibilities.

• H5C_decr__off: Don't attempt to decrease the size of the cache automatically.
  When this increment mode is selected, the remaining fields in the cache size decrease section are ignored.
• H5C_decr__threshold: Attempt to decrease the size of the cache whenever the average hit rate over the last epoch rises above the value supplied in the upper_hr_threshold field.
• H5C_decr__age_out: At the end of each epoch, search the cache for entries that have not been accessed for at least the number of epochs specified in the epochs_before_eviction field, and evict these entries. Conceptually, the maximum cache size is then decreased to match the new actual cache size. However, this reduction may be modified by the min_size, the max_decrement, and/or the empty_reserve.
• H5C_decr__age_out_with_threshold: Same as age_out, but we only attempt to reduce the cache size when the hit rate observed over the last epoch exceeds the value provided in the upper_hr_threshold field.

Note that you must set decr_mode to H5C_decr__off if you disable metadata cache entry evictions.

upper_hr_threshold

Upper hit rate threshold. The use of this field varies according to the current decr_mode :

H5C_decr__off or H5C_decr__age_out: The value of this field is ignored.

• H5C_decr__threshold: If the hit rate exceeds this threshold in any epoch, attempt to decrement the cache size by size_decrement.
  Note that cache size may not be decremented below min_size.
  Note also that if the upper_threshold is 1.0, the cache size
  will never be reduced.

• H5C_decr__age_out_with_threshold: If the hit rate exceeds this threshold in any epoch, attempt to reduce the cache size by evicting entries that have not been accessed for more than the specified number of epochs.

decrement

This field is only used when the decr_mode is H5C_decr__threshold.

The field is a double containing the multiplier used to derive the new cache size from the old if a cache size decrement is triggered. The decrement must be in the range 0.0 (in which case the cache will try to contract to its minimum size) to 1.0 (in which case the cache will never shrink).

apply_max_decrement

Boolean flag used to determine whether decrements in cache size are to be limited by the max_decrement field.

max_decrement

Maximum number of bytes by which the cache size can be decreased in a single re-size. Note that decrements may also be restricted by the min_size of the cache, and (in age out modes) by the empty_reserve field.

epochs_before_eviction

Integer field used in H5C_decr__age_out and H5C_decr__age_out_with_threshold decrement modes.

This field contains the number of epochs an entry must remain unaccessed before it is evicted in an attempt to reduce the cache size. If applicable, this field must lie in the range [1, H5C__MAX_EPOCH_MARKERS].

apply_empty_reserve

Boolean field controlling whether the empty_reserve field is to be used in computing the new cache size when the decr_mode is H5C_decr__age_out or H5C_decr__age_out_with_threshold.

empty_reserve

To avoid a constant racheting down of cache size by small amounts in the H5C_decr__age_out and H5C_decr__age_out_with_threshold modes, this field allows one to require that any cache size reductions leave the specified fraction of unused space in the cache.

The value of this field must be in the range [0.0, 1.0]. I would expect typical values to be in the range of 0.01 to 0.1.

Parallel Configuration Fields

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, which can result in inconsistencies between the caches.

PHDF5 uses several strategies to prevent such inconsistencies in metadata, all of which use the fact that the same stream of dirty metadata is seen by all processes for purposes of synchronization. This is done by having each process count the number of bytes of dirty metadata generated, and then running a "sync point" whenever this count exceeds a user specified threshold (see dirty_bytes_threshold below).

The current metadata write strategy is indicated by the metadata_write_strategy field. The possible values of this field, along with the associated metadata write strategies are discussed below.

dirty_bytes_threshold

Threshold of dirty byte creation used to synchronize updates between caches. (See above for outline and motivation.)

This value MUST be consistent across all processes accessing the file. This field is ignored unless HDF5 has been compiled for parallel.

metadata_write_strategy

Integer field containing a code indicating the desired metadata write strategy. The valid values of this field are enumerated and discussed below:

• H5AC_METADATA_WRITE_STRATEGY__PROCESS_0_ONLY
  When metadata_write_strategy is set to this value, only process zero is allowed to write dirty metadata to disk. All other processes must retain dirty metadata until they are informed at a sync point that the dirty metadata in question has been written to disk.
  When the sync point is reached (or when there is a user generated flush), process zero flushes sufficient entries to bring it into complience with its min clean size (or flushes all dirty entries in the case of a user generated flush), broad casts the list of entries just cleaned to all the other processes, and then exits the sync point.
  Upon receipt of the broadcast, the other processes mark the indicated entries as clean, and leave the sync point as well.

• H5AC_METADATA_WRITE_STRATEGY__DISTRIBUTED
  In the distributed metadata write strategy, process zero still makes the decisions as to what entries should be flushed, but the actual flushes are distributed across the processes in the computation to the extent possible.
  In this strategy, when a sync point is triggered (either by dirty metadata creation or manual flush), all processes enter a barrier.
  On the other side of the barrier, process 0 constructs an ordered list of the entries to be flushed, and then broadcasts this list to the caches in all the processes.
  All processes then scan the list of entries to be flushed, flushing some, and marking the rest as clean. The algorithm for this purpose ensures that each entry in the list is flushed exactly once, and all are marked clean in each cache.
  Note that in the case of a flush of the cache, no message passing is necessary, as all processes have the same list of dirty entries, and all of these entries must be flushed. Thus in this case it is sufficient for each process to sort its list of dirty entries after leaving the initial barrier, and use this list as if it had been received from process zero.
  To avoid possible messages from the past/future, all caches must wait until all caches are done before leaving the sync point.