/* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * Copyright by the Board of Trustees of the University of Illinois. * * All rights reserved. * * * * This file is part of HDF5. The full HDF5 copyright notice, including * * terms governing use, modification, and redistribution, is contained in * * the files COPYING and Copyright.html. COPYING can be found at the root * * of the source code distribution tree; Copyright.html can be found at the * * root level of an installed copy of the electronic HDF5 document set and * * is linked from the top-level documents page. It can also be found at * * http://hdf.ncsa.uiuc.edu/HDF5/doc/Copyright.html. If you do not have * * access to either file, you may request a copy from hdfhelp@ncsa.uiuc.edu. * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */ /* * Programmer: John Mainzer -- 10/12/04 * * Purpose: This file contains declarations which are normally visible * only within the H5C package (just H5C.c at present). * * Source files outside the H5C package should include * H5Cprivate.h instead. * * The one exception to this rule is test/cache.c. The test * code is easier to write if it can look at the cache's * internal data structures. Indeed, this is the main * reason why this file was created. */ #ifndef H5C_PACKAGE #error "Do not include this file outside the H5HL package!" #endif #ifndef _H5Cpkg_H #define _H5Cpkg_H /* Get package's private header */ #include "H5Cprivate.h" /* Get needed headers */ #include "H5SLprivate.h" /* Skip lists */ #define H5C__HASH_TABLE_LEN (64 * 1024) /* must be a power of 2 */ /**************************************************************************** * * structure H5C_t * * Catchall structure for all variables specific to an instance of the cache. * * While the individual fields of the structure are discussed below, the * following overview may be helpful. * * Entries in the cache are stored in an instance of H5TB_TREE, indexed on * the entry's disk address. While the H5TB_TREE is less efficient than * hash table, it keeps the entries in address sorted order. As flushes * in parallel mode are more efficient if they are issued in increasing * address order, this is a significant benefit. Also the H5TB_TREE code * was readily available, which reduced development time. * * While the cache was designed with multiple replacement policies in mind, * at present only a modified form of LRU is supported. * * JRM - 4/26/04 * * Profiling has indicated that searches in the instance of H5TB_TREE are * too expensive. To deal with this issue, I have augmented the cache * with a hash table in which all entries will be stored. Given the * advantages of flushing entries in increasing address order, the TBBT * is retained, but only dirty entries are stored in it. At least for * now, we will leave entries in the TBBT after they are flushed. * * Note that index_size and index_len now refer to the total size of * and number of entries in the hash table. * * JRM - 7/19/04 * * magic: Unsigned 32 bit integer always set to H5C__H5C_T_MAGIC. This * field is used to validate pointers to instances of H5C_t. * * max_type_id: Integer field containing the maximum type id number assigned * to a type of entry in the cache. All type ids from 0 to * max_type_id inclusive must be defined. The names of the * types are stored in the type_name_table discussed below, and * indexed by the ids. * * type_name_table_ptr: Pointer to an array of pointer to char of length * max_type_id + 1. The strings pointed to by the entries * in the array are the names of the entry types associated * with the indexing type IDs. * * max_cache_size: Nominal maximum number of bytes that may be stored in the * cache. This value should be viewed as a soft limit, as the * cache can exceed this value under the following circumstances: * * a) All entries in the cache are protected, and the cache is * asked to insert a new entry. In this case the new entry * will be created. If this causes the cache to exceed * max_cache_size, it will do so. The cache will attempt * to reduce its size as entries are unprotected. * * b) When running in parallel mode, the cache may not be * permitted to flush a dirty entry in response to a read. * If there are no clean entries available to evict, the * cache will exceed its maximum size. Again the cache * will attempt to reduce its size to the max_cache_size * limit on the next cache write. * * min_clean_size: Nominal minimum number of clean bytes in the cache. * The cache attempts to maintain this number of bytes of * clean data so as to avoid case b) above. Again, this is * a soft limit. * * * In addition to the call back functions required for each entry, the * cache requires the following call back functions for this instance of * the cache as a whole: * * check_write_permitted: In certain applications, the cache may not * be allowed to write to disk at certain time. If specified, * the check_write_permitted function is used to determine if * a write is permissible at any given point in time. * * If no such function is specified (i.e. this field is NULL), * the cache will presume that writes are always permissable. * * * The cache requires an index to facilitate searching for entries. The * following fields support that index. * * index_len: Number of entries currently in the hash table used to index * the cache. * * index_size: Number of bytes of cache entries currently stored in the * hash table used to index the cache. * * This value should not be mistaken for footprint of the * cache in memory. The average cache entry is small, and * the cache has a considerable overhead. Multiplying the * index_size by two should yield a conservative estimate * of the cache's memory footprint. * * index: Array of pointer to H5C_cache_entry_t of size * H5C__HASH_TABLE_LEN. At present, this value is a power * of two, not the usual prime number. * * I hope that the variable size of cache elements, the large * hash table size, and the way in which HDF5 allocates space * will combine to avoid problems with periodicity. If so, we * can use a trivial hash function (a bit-and and a 3 bit left * shift) with some small savings. * * If not, it will become evident in the statistics. Changing * to the usual prime number length hash table will require * changing the H5C__HASH_FCN macro and the deletion of the * H5C__HASH_MASK #define. No other changes should be required. * * * When we flush the cache, we need to write entries out in increasing * address order. An instance of a skip list is used to store dirty entries in * sorted order. Whether it is cheaper to sort the dirty entries as needed, * or to maintain the list is an open question. At a guess, it depends * on how frequently the cache is flushed. We will see how it goes. * * For now at least, I will not remove dirty entries from the list as they * are flushed. * * slist_len: Number of entries currently in the skip list * used to maintain a sorted list of dirty entries in the * cache. * * slist_size: Number of bytes of cache entries currently stored in the * skip list used to maintain a sorted list of * dirty entries in the cache. * * slist_ptr: pointer to the instance of H5SL_t used maintain a sorted * list of dirty entries in the cache. This sorted list has * two uses: * * a) It allows us to flush dirty entries in increasing address * order, which results in significant savings. * * b) It facilitates checking for adjacent dirty entries when * attempting to evict entries from the cache. While we * don't use this at present, I hope that this will allow * some optimizations when I get to it. * * * When a cache entry is protected, it must be removed from the LRU * list(s) as it cannot be either flushed or evicted until it is unprotected. * The following fields are used to implement the protected list (pl). * * pl_len: Number of entries currently residing on the protected list. * * pl_size: Number of bytes of cache entries currently residing on the * protected list. * * pl_head_ptr: Pointer to the head of the doubly linked list of protected * entries. Note that cache entries on this list are linked * by their next and prev fields. * * This field is NULL if the list is empty. * * pl_tail_ptr: Pointer to the tail of the doubly linked list of protected * entries. Note that cache entries on this list are linked * by their next and prev fields. * * This field is NULL if the list is empty. * * * The cache must have a replacement policy, and the fields supporting this * policy must be accessible from this structure. * * While there has been interest in several replacement policies for * this cache, the initial development schedule is tight. Thus I have * elected to support only a modified LRU policy for the first cut. * * To further simplify matters, I have simply included the fields needed * by the modified LRU in this structure. When and if we add support for * other policies, it will probably be easiest to just add the necessary * fields to this structure as well -- we only create one instance of this * structure per file, so the overhead is not excessive. * * * Fields supporting the modified LRU policy: * * See most any OS text for a discussion of the LRU replacement policy. * * When operating in parallel mode, we must ensure that a read does not * cause a write. If it does, the process will hang, as the write will * be collective and the other processes will not know to participate. * * To deal with this issue, I have modified the usual LRU policy by adding * clean and dirty LRU lists to the usual LRU list. * * The clean LRU list is simply the regular LRU list with all dirty cache * entries removed. * * Similarly, the dirty LRU list is the regular LRU list with all the clean * cache entries removed. * * When reading in parallel mode, we evict from the clean LRU list only. * This implies that we must try to ensure that the clean LRU list is * reasonably well stocked at all times. * * We attempt to do this by trying to flush enough entries on each write * to keep the cLRU_list_size >= min_clean_size. * * Even if we start with a completely clean cache, a sequence of protects * without unprotects can empty the clean LRU list. In this case, the * cache must grow temporarily. At the next write, we will attempt to * evict enough entries to reduce index_size to less than max_cache_size. * While this will usually be possible, all bets are off if enough entries * are protected. * * Discussions of the individual fields used by the modified LRU replacement * policy follow: * * LRU_list_len: Number of cache entries currently on the LRU list. * * Observe that LRU_list_len + pl_len must always equal * index_len. * * LRU_list_size: Number of bytes of cache entries currently residing on the * LRU list. * * Observe that LRU_list_size + pl_size must always equal * index_size. * * LRU_head_ptr: Pointer to the head of the doubly linked LRU list. Cache * entries on this list are linked by their next and prev fields. * * This field is NULL if the list is empty. * * LRU_tail_ptr: Pointer to the tail of the doubly linked LRU list. Cache * entries on this list are linked by their next and prev fields. * * This field is NULL if the list is empty. * * cLRU_list_len: Number of cache entries currently on the clean LRU list. * * Observe that cLRU_list_len + dLRU_list_len must always * equal LRU_list_len. * * cLRU_list_size: Number of bytes of cache entries currently residing on * the clean LRU list. * * Observe that cLRU_list_size + dLRU_list_size must always * equal LRU_list_size. * * cLRU_head_ptr: Pointer to the head of the doubly linked clean LRU list. * Cache entries on this list are linked by their aux_next and * aux_prev fields. * * This field is NULL if the list is empty. * * cLRU_tail_ptr: Pointer to the tail of the doubly linked clean LRU list. * Cache entries on this list are linked by their aux_next and * aux_prev fields. * * This field is NULL if the list is empty. * * dLRU_list_len: Number of cache entries currently on the dirty LRU list. * * Observe that cLRU_list_len + dLRU_list_len must always * equal LRU_list_len. * * dLRU_list_size: Number of cache entries currently on the dirty LRU list. * * Observe that cLRU_list_len + dLRU_list_len must always * equal LRU_list_len. * * dLRU_head_ptr: Pointer to the head of the doubly linked dirty LRU list. * Cache entries on this list are linked by their aux_next and * aux_prev fields. * * This field is NULL if the list is empty. * * dLRU_tail_ptr: Pointer to the tail of the doubly linked dirty LRU list. * Cache entries on this list are linked by their aux_next and * aux_prev fields. * * This field is NULL if the list is empty. * * * Automatic cache size adjustment: * * While the default cache size is adequate for most cases, we can run into * cases where the default is too small. Ideally, we will let the user * adjust the cache size as required. However, this is not possible in all * cases. Thus I have added automatic cache size adjustment code. * * The configuration for the automatic cache size adjustment is stored in * the structure described below: * * size_increase_possible: Depending on the configuration data given * in the resize_ctl field, it may or may not be possible * to increase the size of the cache. Rather than test for * all the ways this can happen, we simply set this flag when * we receive a new configuration. * * size_decrease_possible: Depending on the configuration data given * in the resize_ctl field, it may or may not be possible * to decrease the size of the cache. Rather than test for * all the ways this can happen, we simply set this flag when * we receive a new configuration. * * cache_full: Boolean flag used to keep track of whether the cache is * full, so we can refrain from increasing the size of a * cache which hasn't used up the space alotted to it. * * The field is initialized to FALSE, and then set to TRUE * whenever we attempt to make space in the cache. * * resize_enabled: This is another convenience flag which is set whenever * a new set of values for resize_ctl are provided. Very * simply, * * resize_enabled = size_increase_possible || * size_decrease_possible; * * size_decreased: Boolean flag set to TRUE whenever the maximun cache * size is decreased. The flag triggers a call to * H5C_make_space_in_cache() on the next call to H5C_protect(). * * resize_ctl: Instance of H5C_auto_size_ctl_t containing configuration * data for automatic cache resizing. * * epoch_markers_active: Integer field containing the number of epoch * markers currently in use in the LRU list. This value * must be in the range [0, H5C__MAX_EPOCH_MARKERS - 1]. * * epoch_marker_active: Array of boolean of length H5C__MAX_EPOCH_MARKERS. * This array is used to track which epoch markers are currently * in use. * * epoch_marker_ringbuf: Array of int of length H5C__MAX_EPOCH_MARKERS + 1. * * To manage the epoch marker cache entries, it is necessary * to track their order in the LRU list. This is done with * epoch_marker_ringbuf. When markers are inserted at the * head of the LRU list, the index of the marker in the * epoch_markers array is inserted at the tail of the ring * buffer. When it becomes the epoch_marker_active'th marker * in the LRU list, it will have worked its way to the head * of the ring buffer as well. This allows us to remove it * without scanning the LRU list if such is required. * * epoch_marker_ringbuf_first: Integer field containing the index of the * first entry in the ring buffer. * * epoch_marker_ringbuf_last: Integer field containing the index of the * last entry in the ring buffer. * * epoch_marker_ringbuf_size: Integer field containing the number of entries * in the ring buffer. * * epoch_markers: Array of instances of H5C_cache_entry_t of length * H5C__MAX_EPOCH_MARKERS. The entries are used as markers * in the LRU list to identify cache entries that haven't * been accessed for some (small) specified number of * epochs. These entries (if any) can then be evicted and * the cache size reduced -- ideally without evicting any * of the current working set. Needless to say, the epoch * length and the number of epochs before an unused entry * must be chosen so that all, or almost all, the working * set will be accessed before the limit. * * Epoch markers only appear in the LRU list, never in * the index or slist. While they are of type * H5C__EPOCH_MARKER_TYPE, and have associated class * functions, these functions should never be called. * * The addr fields of these instances of H5C_cache_entry_t * are set to the index of the instance in the epoch_markers * array, the size is set to 0, and the type field points * to the constant structure epoch_marker_class defined * in H5C.c. The next and prev fields are used as usual * to link the entry into the LRU list. * * All other fields are unused. * * * Cache hit rate collection fields: * * We supply the current cache hit rate on request, so we must keep a * simple cache hit rate computation regardless of whether statistics * collection is enabled. The following fields support this capability. * * cache_hits: Number of cache hits since the last time the cache hit * rate statistics were reset. Note that when automatic cache * re-sizing is enabled, this field will be reset every automatic * resize epoch. * * cache_accesses: Number of times the cache has been accessed while * since the last since the last time the cache hit rate statistics * were reset. Note that when automatic cache re-sizing is enabled, * this field will be reset every automatic resize epoch. * * * Statistics collection fields: * * When enabled, these fields are used to collect statistics as described * below. The first set are collected only when H5C_COLLECT_CACHE_STATS * is true. * * hits: Array of int64 of length H5C__MAX_NUM_TYPE_IDS + 1. The cells * are used to record the number of times an entry with type id * equal to the array index has been in cache when requested in * the current epoch. * * misses: Array of int64 of length H5C__MAX_NUM_TYPE_IDS + 1. The cells * are used to record the number of times an entry with type id * equal to the array index has not been in cache when * requested in the current epoch. * * insertions: Array of int64 of length H5C__MAX_NUM_TYPE_IDS + 1. The cells * are used to record the number of times an entry with type * id equal to the array index has been inserted into the * cache in the current epoch. * * clears: Array of int64 of length H5C__MAX_NUM_TYPE_IDS + 1. The cells * are used to record the number of times an entry with type * id equal to the array index has been cleared in the current * epoch. * * flushes: Array of int64 of length H5C__MAX_NUM_TYPE_IDS + 1. The cells * are used to record the number of times an entry with type id * equal to the array index has been written to disk in the * current epoch. * * evictions: Array of int64 of length H5C__MAX_NUM_TYPE_IDS + 1. The cells * are used to record the number of times an entry with type id * equal to the array index has been evicted from the cache in * the current epoch. * * renames: Array of int64 of length H5C__MAX_NUM_TYPE_IDS + 1. The cells * are used to record the number of times an entry with type * id equal to the array index has been renamed in the current * epoch. * * total_ht_insertions: Number of times entries have been inserted into the * hash table in the current epoch. * * total_ht_deletions: Number of times entries have been deleted from the * hash table in the current epoch. * * successful_ht_searches: int64 containing the total number of successful * searches of the hash table in the current epoch. * * total_successful_ht_search_depth: int64 containing the total number of * entries other than the targets examined in successful * searches of the hash table in the current epoch. * * failed_ht_searches: int64 containing the total number of unsuccessful * searches of the hash table in the current epoch. * * total_failed_ht_search_depth: int64 containing the total number of * entries examined in unsuccessful searches of the hash * table in the current epoch. * * max_index_len: Largest value attained by the index_len field in the * current epoch. * * max_index_size: Largest value attained by the index_size field in the * current epoch. * * max_slist_len: Largest value attained by the slist_len field in the * current epoch. * * max_slist_size: Largest value attained by the slist_size field in the * current epoch. * * max_pl_len: Largest value attained by the pl_len field in the * current epoch. * * max_pl_size: Largest value attained by the pl_size field in the * current epoch. * * The remaining stats are collected only when both H5C_COLLECT_CACHE_STATS * and H5C_COLLECT_CACHE_ENTRY_STATS are true. * * max_accesses: Array of int32 of length H5C__MAX_NUM_TYPE_IDS + 1. The cells * are used to record the maximum number of times any single * entry with type id equal to the array index has been * accessed in the current epoch. * * min_accesses: Array of int32 of length H5C__MAX_NUM_TYPE_IDS + 1. The cells * are used to record the minimum number of times any single * entry with type id equal to the array index has been * accessed in the current epoch. * * max_clears: Array of int32 of length H5C__MAX_NUM_TYPE_IDS + 1. The cells * are used to record the maximum number of times any single * entry with type id equal to the array index has been cleared * in the current epoch. * * max_flushes: Array of int32 of length H5C__MAX_NUM_TYPE_IDS + 1. The cells * are used to record the maximum number of times any single * entry with type id equal to the array index has been * flushed in the current epoch. * * max_size: Array of size_t of length H5C__MAX_NUM_TYPE_IDS + 1. The cells * are used to record the maximum size of any single entry * with type id equal to the array index that has resided in * the cache in the current epoch. * * * Fields supporting testing: * * For test purposes, it is useful to turn off some asserts and sanity * checks. The following flags support this. * * skip_file_checks: Boolean flag used to skip sanity checks on file * parameters passed to the cache. In the test bed, there * is no reason to have a file open, as the cache proper * just passes these parameters through without using them. * * When this flag is set, all sanity checks on the file * parameters are skipped. The field defaults to FALSE. * * skip_dxpl_id_checks: Boolean flag used to skip sanity checks on the * dxpl_id parameters passed to the cache. These are not * used directly by the cache, so skipping the checks * simplifies the test bed. * * When this flag is set, all sanity checks on the dxpl_id * parameters are skipped. The field defaults to FALSE. * ****************************************************************************/ #define H5C__H5C_T_MAGIC 0x005CAC0E #define H5C__MAX_NUM_TYPE_IDS 12 struct H5C_t { uint32_t magic; int32_t max_type_id; const char * (* type_name_table_ptr); size_t max_cache_size; size_t min_clean_size; H5C_write_permitted_func_t check_write_permitted; int32_t index_len; size_t index_size; H5C_cache_entry_t * (index[H5C__HASH_TABLE_LEN]); int32_t slist_len; size_t slist_size; H5SL_t * slist_ptr; int32_t pl_len; size_t pl_size; H5C_cache_entry_t * pl_head_ptr; H5C_cache_entry_t * pl_tail_ptr; int32_t LRU_list_len; size_t LRU_list_size; H5C_cache_entry_t * LRU_head_ptr; H5C_cache_entry_t * LRU_tail_ptr; int32_t cLRU_list_len; size_t cLRU_list_size; H5C_cache_entry_t * cLRU_head_ptr; H5C_cache_entry_t * cLRU_tail_ptr; int32_t dLRU_list_len; size_t dLRU_list_size; H5C_cache_entry_t * dLRU_head_ptr; H5C_cache_entry_t * dLRU_tail_ptr; hbool_t size_increase_possible; hbool_t size_decrease_possible; hbool_t resize_enabled; hbool_t cache_full; hbool_t size_decreased; H5C_auto_size_ctl_t resize_ctl; int32_t epoch_markers_active; hbool_t epoch_marker_active[H5C__MAX_EPOCH_MARKERS]; int32_t epoch_marker_ringbuf[H5C__MAX_EPOCH_MARKERS+1]; int32_t epoch_marker_ringbuf_first; int32_t epoch_marker_ringbuf_last; int32_t epoch_marker_ringbuf_size; H5C_cache_entry_t epoch_markers[H5C__MAX_EPOCH_MARKERS]; int64_t cache_hits; int64_t cache_accesses; #if H5C_COLLECT_CACHE_STATS /* stats fields */ int64_t hits[H5C__MAX_NUM_TYPE_IDS + 1]; int64_t misses[H5C__MAX_NUM_TYPE_IDS + 1]; int64_t insertions[H5C__MAX_NUM_TYPE_IDS + 1]; int64_t clears[H5C__MAX_NUM_TYPE_IDS + 1]; int64_t flushes[H5C__MAX_NUM_TYPE_IDS + 1]; int64_t evictions[H5C__MAX_NUM_TYPE_IDS + 1]; int64_t renames[H5C__MAX_NUM_TYPE_IDS + 1]; int64_t total_ht_insertions; int64_t total_ht_deletions; int64_t successful_ht_searches; int64_t total_successful_ht_search_depth; int64_t failed_ht_searches; int64_t total_failed_ht_search_depth; int32_t max_index_len; size_t max_index_size; int32_t max_slist_len; size_t max_slist_size; int32_t max_pl_len; size_t max_pl_size; #if H5C_COLLECT_CACHE_ENTRY_STATS int32_t max_accesses[H5C__MAX_NUM_TYPE_IDS + 1]; int32_t min_accesses[H5C__MAX_NUM_TYPE_IDS + 1]; int32_t max_clears[H5C__MAX_NUM_TYPE_IDS + 1]; int32_t max_flushes[H5C__MAX_NUM_TYPE_IDS + 1]; size_t max_size[H5C__MAX_NUM_TYPE_IDS + 1]; #endif /* H5C_COLLECT_CACHE_ENTRY_STATS */ #endif /* H5C_COLLECT_CACHE_STATS */ hbool_t skip_file_checks; hbool_t skip_dxpl_id_checks; }; #endif /* _H5Cpkg_H */