summaryrefslogtreecommitdiffstats
path: root/src
diff options
context:
space:
mode:
authorRobb Matzke <matzke@llnl.gov>1998-09-24 15:51:05 (GMT)
committerRobb Matzke <matzke@llnl.gov>1998-09-24 15:51:05 (GMT)
commit311e4c9ebfbea0be881586df16a8a411bb7fc9f8 (patch)
treee92cf9dde45481da9a6070ce0926a1d279538201 /src
parentf180bc993f45bdb5d3a46abeab9e1378e6f210da (diff)
downloadhdf5-311e4c9ebfbea0be881586df16a8a411bb7fc9f8.zip
hdf5-311e4c9ebfbea0be881586df16a8a411bb7fc9f8.tar.gz
hdf5-311e4c9ebfbea0be881586df16a8a411bb7fc9f8.tar.bz2
[svn-r720] Changes since 19980922
---------------------- ./src/H5F.c ./src/H5Fprivate.h ./src/H5P.c ./src/H5Ppublic.h ./test/chunk.c ./test/dsets.c The number of slots in the raw data cache can be queried or set with H5Pget/set_cache(), which now take an extra argument. The default number of slots is 521 and the default maximum size is 1MB. ./src/H5Fistore.c ./src/H5Fprivate.h Finished optimizations. The cache is now a hash and a linked list instead of an array. The cpu time on my machine for H5F_istore_lock() has been cut by 60% and H5F_istore_unlock() by 35%.
Diffstat (limited to 'src')
-rw-r--r--src/H5Distore.c464
-rw-r--r--src/H5F.c1
-rw-r--r--src/H5Fistore.c464
-rw-r--r--src/H5Fprivate.h9
-rw-r--r--src/H5P.c44
-rw-r--r--src/H5Ppublic.h7
6 files changed, 536 insertions, 453 deletions
diff --git a/src/H5Distore.c b/src/H5Distore.c
index dcd1d35..278bca9 100644
--- a/src/H5Distore.c
+++ b/src/H5Distore.c
@@ -13,6 +13,21 @@
* chunk index to disk address. Each chunk can be compressed
* independently and the chunks may move around in the file as
* their storage requirements change.
+ *
+ * Cache: Disk I/O is performed in units of chunks and H5MF_alloc()
+ * contains code to optionally align chunks on disk block
+ * boundaries for performance.
+ *
+ * The chunk cache is an extendible hash indexed by a function
+ * of storage B-tree address and chunk N-dimensional offset
+ * within the dataset. Collisions are not resolved -- one of
+ * the two chunks competing for the hash slot must be preempted
+ * from the cache. All entries in the hash also participate in
+ * a doubly-linked list and entries are penalized by moving them
+ * toward the front of the list. When a new chunk is about to
+ * be added to the cache the heap is pruned by preempting
+ * entries near the front of the list to make room for the new
+ * entry which is added to the end of the list.
*/
#include <H5private.h>
#include <H5Dprivate.h>
@@ -23,6 +38,31 @@
#include <H5Oprivate.h>
#include <H5Vprivate.h>
+/*
+ * Feature: If this constant is defined then every cache preemption and load
+ * causes a character to be printed on the standard error stream:
+ *
+ * `.': Entry was preempted because it has been completely read or
+ * completely written but not partially read and not partially
+ * written. This is often a good reason for preemption because such
+ * a chunk will be unlikely to be referenced in the near future.
+ *
+ * `:': Entry was preempted because it hasn't been used recently.
+ *
+ * `#': Entry was preempted because another chunk collided with it. This
+ * is usually a relatively bad thing. If there are too many of
+ * these then the number of entries in the cache can be increased.
+ *
+ * c: Entry was preempted because the file is closing.
+ *
+ * w: A chunk read operation was eliminated because the library is
+ * about to write new values to the entire chunk. This is a good
+ * thing, especially on files where the chunk size is the same as
+ * the disk block size, chunks are aligned on disk block boundaries,
+ * and the operating system can also eliminate a read operation.
+ */
+/* #define H5F_ISTORE_DEBUG */
+
/* Interface initialization */
#define PABLO_MASK H5F_istore_mask
static hbool_t interface_initialize_g = FALSE;
@@ -40,6 +80,9 @@ typedef struct H5F_rdcc_ent_t {
size_t chunk_size; /*size of a chunk */
size_t alloc_size; /*amount allocated for the chunk */
uint8 *chunk; /*the unfiltered chunk data */
+ intn idx; /*index in hash table */
+ struct H5F_rdcc_ent_t *next;/*next item in doubly-linked list */
+ struct H5F_rdcc_ent_t *prev;/*previous item in doubly-linked list */
} H5F_rdcc_ent_t;
/* Private prototypes */
@@ -688,9 +731,10 @@ H5F_istore_init (H5F_t *f)
FUNC_ENTER (H5F_istore_init, FAIL);
HDmemset (rdcc, 0, sizeof(H5F_rdcc_t));
- if (f->shared->access_parms->rdcc_nbytes>0) {
- rdcc->nslots = 25; /*some initial number of slots*/
- rdcc->slot = H5MM_calloc (rdcc->nslots*sizeof(H5F_rdcc_ent_t));
+ if (f->shared->access_parms->rdcc_nbytes>0 &&
+ f->shared->access_parms->rdcc_nelmts>0) {
+ rdcc->nslots = f->shared->access_parms->rdcc_nelmts;
+ rdcc->slot = H5MM_calloc (rdcc->nslots*sizeof(H5F_rdcc_ent_t*));
if (NULL==rdcc->slot) {
HRETURN_ERROR (H5E_RESOURCE, H5E_NOSPACE, FAIL,
"memory allocation failed");
@@ -797,7 +841,7 @@ H5F_istore_flush_entry (H5F_t *f, H5F_rdcc_ent_t *ent, hbool_t reset)
f->shared->rdcc.nflushes++;
}
- /* Reset */
+ /* Reset, but do not free or removed from list */
if (reset) {
point_of_no_return = FALSE;
ent->layout = H5O_free(H5O_LAYOUT, ent->layout);
@@ -815,7 +859,8 @@ H5F_istore_flush_entry (H5F_t *f, H5F_rdcc_ent_t *ent, hbool_t reset)
/*
* If we reached the point of no return then we have no choice but to
* reset the entry. This can only happen if RESET is true but the
- * output pipeline failed.
+ * output pipeline failed. Do not free the entry or remove it from the
+ * list.
*/
if (ret_value<0 && point_of_no_return) {
ent->layout = H5O_free(H5O_LAYOUT, ent->layout);
@@ -846,25 +891,17 @@ H5F_istore_flush_entry (H5F_t *f, H5F_rdcc_ent_t *ent, hbool_t reset)
herr_t
H5F_istore_flush (H5F_t *f)
{
- H5F_rdcc_t *rdcc = &(f->shared->rdcc);
- intn i, nerrors=0;
+ H5F_rdcc_t *rdcc = &(f->shared->rdcc);
+ intn nerrors=0;
+ H5F_rdcc_ent_t *ent=NULL;
FUNC_ENTER (H5F_istore_flush, FAIL);
-#ifdef H5F_RDCC_NEW
- for (i=0; i<rdcc->nslots; i++) {
- if (rdcc->slot[i].chunk &&
- H5F_istore_flush_entry(f, rdcc->slot+i, FALSE)<0) {
- nerrors++;
- }
- }
-#else
- for (i=0; i<rdcc->nused; i++) {
- if (H5F_istore_flush_entry (f, rdcc->slot+i, FALSE)<0) {
+ for (ent=rdcc->head; ent; ent=ent->next) {
+ if (H5F_istore_flush_entry(f, ent, FALSE)<0) {
nerrors++;
}
}
-#endif
if (nerrors) {
HRETURN_ERROR (H5E_IO, H5E_CANTFLUSH, FAIL,
"unable to flush one or more raw data chunks");
@@ -890,29 +927,41 @@ H5F_istore_flush (H5F_t *f)
*-------------------------------------------------------------------------
*/
static herr_t
-H5F_istore_preempt (H5F_t *f, intn idx)
+H5F_istore_preempt (H5F_t *f, H5F_rdcc_ent_t *ent)
{
H5F_rdcc_t *rdcc = &(f->shared->rdcc);
- H5F_rdcc_ent_t *ent = rdcc->slot + idx;
FUNC_ENTER (H5F_istore_preempt, FAIL);
-#ifdef H5F_RDCC_NEW
- assert(idx>=0 && idx<rdcc->nslots);
- assert (!ent->locked);
+ assert(f);
+ assert(ent);
+ assert(!ent->locked);
+ assert(ent->idx>=0 && ent->idx<rdcc->nslots);
+ /* Flush */
H5F_istore_flush_entry(f, ent, TRUE);
- rdcc->nbytes -= ent->chunk_size;
-#else
- assert (idx>=0 && idx<rdcc->nused);
- assert (!ent->locked);
- H5F_istore_flush_entry (f, ent, TRUE);
- HDmemmove (rdcc->slot+idx, rdcc->slot+idx+1,
- (rdcc->nused-(idx+1)) * sizeof(H5F_rdcc_ent_t));
- rdcc->nused -= 1;
+ /* Unlink from list */
+ if (ent->prev) {
+ ent->prev->next = ent->next;
+ } else {
+ rdcc->head = ent->next;
+ }
+ if (ent->next) {
+ ent->next->prev = ent->prev;
+ } else {
+ rdcc->tail = ent->prev;
+ }
+ ent->prev = ent->next = NULL;
+
+ /* Remove from cache */
+ rdcc->slot[ent->idx] = NULL;
+ ent->idx = -1;
rdcc->nbytes -= ent->chunk_size;
-#endif
+ --rdcc->nused;
+
+ /* Free */
+ H5MM_xfree(ent);
FUNC_LEAVE (SUCCEED);
}
@@ -938,25 +987,22 @@ H5F_istore_preempt (H5F_t *f, intn idx)
herr_t
H5F_istore_dest (H5F_t *f)
{
- H5F_rdcc_t *rdcc = &(f->shared->rdcc);
- intn i, nerrors=0;
+ H5F_rdcc_t *rdcc = &(f->shared->rdcc);
+ intn nerrors=0;
+ H5F_rdcc_ent_t *ent=NULL, *next=NULL;
FUNC_ENTER (H5F_istore_dest, FAIL);
-#ifdef H5F_RDCC_NEW
- for (i=0; i<rdcc->nslots; i++) {
- if (rdcc->slot[i].chunk &&
- H5F_istore_preempt(f, i)<0) {
- nerrors++;
- }
- }
-#else
- for (i=rdcc->nused-1; i>=0; --i) {
- if (H5F_istore_preempt(f, i)<0) {
+ for (ent=rdcc->head; ent; ent=next) {
+#ifdef H5F_ISTORE_DEBUG
+ fputc('c', stderr);
+ fflush(stderr);
+#endif
+ next = ent->next;
+ if (H5F_istore_preempt(f, ent)<0) {
nerrors++;
}
}
-#endif
if (nerrors) {
HRETURN_ERROR (H5E_IO, H5E_CANTFLUSH, FAIL,
"unable to flush one or more raw data chunks");
@@ -989,71 +1035,89 @@ H5F_istore_dest (H5F_t *f)
static herr_t
H5F_istore_prune (H5F_t *f, size_t size)
{
-#ifdef H5F_RDCC_NEW
- intn i, nerrors=0;
- static intn place=0;
- H5F_rdcc_t *rdcc = &(f->shared->rdcc);
- H5F_rdcc_ent_t *ent = NULL;
- size_t total = f->shared->access_parms->rdcc_nbytes;
-#else
- intn i, meth0, meth1, nerrors=0;
+ intn i, j, nerrors=0;
H5F_rdcc_t *rdcc = &(f->shared->rdcc);
- H5F_rdcc_ent_t *ent0, *ent1;
- double w0 = f->shared->access_parms->rdcc_w0;
size_t total = f->shared->access_parms->rdcc_nbytes;
-#endif
-
+ const int nmeth=2; /*number of methods */
+ intn w[1]; /*weighting as an interval */
+ H5F_rdcc_ent_t *p[2], *cur; /*list pointers */
+ H5F_rdcc_ent_t *n[2]; /*list next pointers */
+
FUNC_ENTER (H5F_istore_prune, FAIL);
-#ifdef H5F_RDCC_NEW
- for (i=0; rdcc->nbytes+size>total && i<rdcc->nslots; i++, place++) {
- if (place>=rdcc->nslots) place = 0;
- ent = rdcc->slot+place;
- if (ent->chunk && !ent->locked) {
- if (H5F_istore_preempt(f, place)<0) nerrors++;
- }
- }
-#else
/*
- * We have two pointers that slide down the cache beginning at the least
- * recently used entry. The distance between the pointers represents the
- * relative weight. A weight of 50% for the first pointer means that the
- * second pointer is half the cache length behind the first pointer.
+ * Preemption is accomplished by having multiple pointers (currently two)
+ * slide down the list beginning at the head. Pointer p(N+1) will start
+ * traversing the list when pointer pN reaches wN percent of the original
+ * list. In other words, preemption method N gets to consider entries in
+ * approximate least recently used order w0 percent before method N+1
+ * where 100% means tha method N will run to completion before method N+1
+ * begins. The pointers participating in the list traversal are each
+ * given a chance at preemption before any of the pointers are advanced.
*/
- meth0 = rdcc->nused;
- meth1 = rdcc->nused * (1.0+w0);
- for (i=MAX(meth0, meth1)-1;
- rdcc->nbytes+size>total && i>=0;
- --i, --meth0, --meth1) {
-
- ent0 = rdcc->slot+meth0; /*might be a bad pointer!*/
- ent1 = rdcc->slot+meth1; /*might be a bad pointer!*/
+ w[0] = rdcc->nused * f->shared->access_parms->rdcc_w0;
+ p[0] = rdcc->head;
+ p[1] = NULL;
+
+ while ((p[0] || p[1]) && rdcc->nbytes+size>total) {
+
+ /* Introduce new pointers */
+ for (i=0; i<nmeth-1; i++) if (0==w[i]) p[i+1] = rdcc->head;
- if (meth0>=0 && meth0<rdcc->nused && !ent0->locked &&
- (0==ent0->rd_count || ent0->chunk_size==ent0->rd_count) &&
- (0==ent0->wr_count || ent0->chunk_size==ent0->wr_count)) {
- /*
- * Method 0: Preempt entries that have a zero rd_count. If the
- * application is accessing a dataset with a set of
- * non-overlapping partial I/O requests then chunks with a zero
- * rd_count will probably not be accessed in the near future.
- */
- if (H5F_istore_preempt (f, meth0)<0) nerrors++;
-
- } else if (meth1>=0 && meth1<rdcc->nused && !ent1->locked) {
- /*
- * Method 1: Discard the least recently used members from the
- * cache. This is a catch-all.
- */
- if (H5F_istore_preempt (f, meth1)<0) nerrors++;
+ /* Compute next value for each pointer */
+ for (i=0; i<nmeth; i++) n[i] = p[i] ? p[i]->next : NULL;
+
+ /* Give each method a chance */
+ for (i=0; i<nmeth && rdcc->nbytes+size>total; i++) {
+ if (0==i && p[0] && !p[0]->locked &&
+ ((0==p[0]->rd_count && 0==p[0]->wr_count) ||
+ (0==p[0]->rd_count && p[0]->chunk_size==p[0]->wr_count) ||
+ (p[0]->chunk_size==p[0]->rd_count && 0==p[0]->wr_count))) {
+ /*
+ * Method 0: Preempt entries that have been completely written
+ * and/or completely read but not entries that are partially
+ * written or partially read.
+ */
+ cur = p[0];
+#ifdef H5F_ISTORE_DEBUG
+ putc('.', stderr);
+ fflush(stderr);
+#endif
+
+ } else if (1==i && p[1] && !p[1]->locked) {
+ /*
+ * Method 1: Preempt the entry without regard to
+ * considerations other than being locked. This is the last
+ * resort preemption.
+ */
+ cur = p[1];
+#ifdef H5F_ISTORE_DEBUG
+ putc(':', stderr);
+ fflush(stderr);
+#endif
+
+ } else {
+ /* Nothing to preempt at this point */
+ cur= NULL;
+ }
+
+ if (cur) {
+ if (H5F_istore_preempt(f, cur)<0) nerrors++;
+ for (j=0; j<nmeth; j++) {
+ if (p[j]==cur) p[j] = NULL;
+ }
+ }
}
+
+ /* Advance pointers */
+ for (i=0; i<nmeth; i++) p[i] = n[i];
+ for (i=0; i<nmeth-1; i++) w[i] -= 1;
}
-#endif
+
if (nerrors) {
HRETURN_ERROR (H5E_IO, H5E_CANTFLUSH, FAIL,
"unable to preempt one or more raw data cache entry");
}
-
FUNC_LEAVE (SUCCEED);
}
@@ -1092,12 +1156,11 @@ H5F_istore_lock (H5F_t *f, const H5O_layout_t *layout,
const H5O_pline_t *pline, const hssize_t offset[],
hbool_t relax, intn *idx_hint/*in,out*/)
{
-#ifdef H5F_RDCC_NEW
- uintn idx;
-#endif
+ uintn idx; /*hash index number */
+ hbool_t found = FALSE; /*already in cache? */
H5F_rdcc_t *rdcc = &(f->shared->rdcc);/*raw data chunk cache*/
H5F_rdcc_ent_t *ent = NULL; /*cache entry */
- intn i, j, found = -1; /*counters */
+ intn i; /*counters */
H5F_istore_ud1_t udata; /*B-tree pass-through */
size_t chunk_size=0; /*size of a chunk */
size_t chunk_alloc=0; /*allocated chunk size */
@@ -1107,61 +1170,41 @@ H5F_istore_lock (H5F_t *f, const H5O_layout_t *layout,
FUNC_ENTER (H5F_istore_lock, NULL);
-#ifdef H5F_RDCC_NEW
if (rdcc->nslots>0) {
idx = layout->addr.offset;
- for (i=0; i<layout->ndims; i++) idx ^= offset[i];
+ for (i=0; i<layout->ndims; i++) idx = (idx<<1) ^ offset[i];
idx %= rdcc->nslots;
- ent = rdcc->slot + idx;
+ ent = rdcc->slot[idx];
- if (ent->chunk &&
+ if (ent &&
layout->ndims==ent->layout->ndims &&
H5F_addr_eq(&(layout->addr), &(ent->layout->addr))) {
- for (i=0, found=idx; i<ent->layout->ndims; i++) {
+ for (i=0, found=TRUE; i<ent->layout->ndims; i++) {
if (offset[i]!=ent->offset[i]) {
- found = -1;
+ found = FALSE;
break;
}
}
}
}
-#else
- /* First use the hint because that's O(1) */
- if (idx_hint && *idx_hint>=0 && *idx_hint<rdcc->nused) {
- ent = rdcc->slot + *idx_hint;
- if (layout->ndims==ent->layout->ndims &&
- H5F_addr_eq(&(layout->addr), &(ent->layout->addr))) {
- for (i=0, found=*idx_hint; found>=0 && i<ent->layout->ndims; i++) {
- if (offset[i]!=ent->offset[i]) found = -1;
- }
- }
- }
-
- /* If the hint is wrong then search the cache, O(n) */
- for (i=0; found<0 && i<rdcc->nused; i++) {
- ent = rdcc->slot + i;
- if (layout->ndims==ent->layout->ndims &&
- H5F_addr_eq(&(layout->addr), &(ent->layout->addr))) {
- for (j=0, found=i; found>=0 && j<ent->layout->ndims; j++) {
- if (offset[j]!=ent->offset[j]) found = -1;
- }
- }
- }
-#endif
- if (found>=0) {
+ if (found) {
/*
* Already in the cache. Count a hit.
*/
rdcc->nhits++;
-
- } else if (found<0 && relax) {
+
+ } else if (!found && relax) {
/*
* Not in the cache, but we're about to overwrite the whole thing
* anyway, so just allocate a buffer for it but don't initialize that
* buffer with the file contents. Count this as a hit instead of a
* miss because we saved ourselves lots of work.
*/
+#ifdef H5F_ISTORE_DEBUG
+ putc('w', stderr);
+ fflush(stderr);
+#endif
rdcc->nhits++;
for (i=0, chunk_size=1; i<layout->ndims; i++) {
chunk_size *= layout->dim[i];
@@ -1215,26 +1258,44 @@ H5F_istore_lock (H5F_t *f, const H5O_layout_t *layout,
rdcc->ninits++;
}
}
-
- assert (found>=0 || chunk_size>0);
-#ifdef H5F_RDCC_NEW
- if (found<0 && rdcc->nslots>0 &&
+ assert (found || chunk_size>0);
+
+ if (!found && rdcc->nslots>0 &&
chunk_size<=f->shared->access_parms->rdcc_nbytes &&
- (NULL==ent->chunk || !ent->locked)) {
+ (!ent || !ent->locked)) {
/*
* Add the chunk to the cache only if the slot is not already locked.
* Preempt enough things from the cache to make room.
*/
+ if (ent) {
+#ifdef H5F_ISTORE_DEBUG
+ putc('#', stderr);
+ fflush(stderr);
+#endif
+#if 0
+ HDfprintf(stderr, "\ncollision %3d %10a {",
+ idx, &(ent->layout->addr));
+ for (i=0; i<layout->ndims; i++) {
+ HDfprintf(stderr, "%s%Zu", i?",":"", ent->offset[i]);
+ }
+ HDfprintf(stderr, "}\n %10a {", &(layout->addr));
+ for (i=0; i<layout->ndims; i++) {
+ HDfprintf(stderr, "%s%Zu", i?",":"", offset[i]);
+ }
+ fprintf(stderr, "}\n");
+#endif
+ if (H5F_istore_preempt(f, ent)<0) {
+ HGOTO_ERROR(H5E_IO, H5E_CANTINIT, NULL,
+ "unable to preempt chunk from cache");
+ }
+ }
if (H5F_istore_prune(f, chunk_size)<0) {
HGOTO_ERROR(H5E_IO, H5E_CANTINIT, NULL,
"unable to preempt chunk(s) from cache");
}
- if (rdcc->slot[idx].chunk &&
- H5F_istore_preempt(f, idx)<0) {
- HGOTO_ERROR(H5E_IO, H5E_CANTINIT, NULL,
- "unable to preempt chunk from cache");
- }
- ent = rdcc->slot + idx;
+
+ /* Create a new entry */
+ ent = H5MM_malloc(sizeof(H5F_rdcc_ent_t));
ent->locked = 0;
ent->dirty = FALSE;
ent->chunk_size = chunk_size;
@@ -1247,83 +1308,67 @@ H5F_istore_lock (H5F_t *f, const H5O_layout_t *layout,
ent->rd_count = chunk_size;
ent->wr_count = chunk_size;
ent->chunk = chunk;
- found = idx;
- }
-#else
- if (found<0 && chunk_size<=f->shared->access_parms->rdcc_nbytes) {
- /*
- * Add the chunk to the beginning of the cache after pruning the cache
- * to make room.
- */
- if (H5F_istore_prune (f, chunk_size)<0) {
- H5E_clear ();
- }
- if (rdcc->nused>=rdcc->nslots) {
- size_t na = MAX (25, 2*rdcc->nslots);
- H5F_rdcc_ent_t *x = H5MM_realloc (rdcc->slot,
- na*sizeof(H5F_rdcc_ent_t));
- if (NULL==x) {
- HGOTO_ERROR (H5E_RESOURCE, H5E_NOSPACE, NULL,
- "memory allocation failed");
- }
- rdcc->nslots = (intn)na;
- rdcc->slot = x;
- }
- HDmemmove (rdcc->slot+1, rdcc->slot,
- rdcc->nused*sizeof(H5F_rdcc_ent_t));
- rdcc->nused++;
+
+ /* Add it to the cache */
+ assert(NULL==rdcc->slot[idx]);
+ rdcc->slot[idx] = ent;
+ ent->idx = idx;
rdcc->nbytes += chunk_size;
- ent = rdcc->slot;
- ent->locked = 0;
- ent->dirty = FALSE;
- ent->chunk_size = chunk_size;
- ent->alloc_size = chunk_alloc;
- ent->layout = H5O_copy (H5O_LAYOUT, layout, NULL);
- ent->pline = H5O_copy (H5O_PLINE, pline, NULL);
- for (i=0; i<layout->ndims; i++) {
- ent->offset[i] = offset[i];
+ rdcc->nused++;
+
+ /* Add it to the linked list */
+ ent->next = NULL;
+ if (rdcc->tail) {
+ rdcc->tail->next = ent;
+ ent->prev = rdcc->tail;
+ rdcc->tail = ent;
+ } else {
+ rdcc->head = rdcc->tail = ent;
+ ent->prev = NULL;
}
- ent->rd_count = chunk_size;
- ent->wr_count = chunk_size;
- ent->chunk = chunk;
- found = 0;
- }
-#endif
- else if (found<0) {
+ found = TRUE;
+
+ } else if (!found) {
/*
* The chunk is larger than the entire cache so we don't cache it.
* This is the reason all those arguments have to be repeated for the
* unlock function.
*/
ent = NULL;
- found = -999;
+ idx = -999;
-#ifndef H5F_RDCC_NEW
- } else if (found>0) {
+ } else if (found) {
/*
- * The chunk is not at the beginning of the cache; move it forward by
- * one slot. This is how we implement the LRU preemption algorithm.
+ * The chunk is not at the beginning of the cache; move it backward
+ * by one slot. This is how we implement the LRU preemption
+ * algorithm.
*/
- H5F_rdcc_ent_t x = rdcc->slot[found];
- rdcc->slot[found] = rdcc->slot[found-1];
- rdcc->slot[found-1] = x;
- ent = rdcc->slot + --found;
-#endif
+ if (ent->next) {
+ if (ent->next->next) {
+ ent->next->next->prev = ent;
+ } else {
+ rdcc->tail = ent;
+ }
+ ent->next->prev = ent->prev;
+ if (ent->prev) {
+ ent->prev->next = ent->next;
+ } else {
+ rdcc->head = ent->next;
+ }
+ ent->prev = ent->next;
+ ent->next = ent->next->next;
+ ent->prev->next = ent;
+ }
}
/* Lock the chunk into the cache */
if (ent) {
assert (!ent->locked);
ent->locked = TRUE;
-#ifndef H5F_RDCC_NEW
- if (idx_hint) *idx_hint = found;
-#endif
chunk = ent->chunk;
}
-#ifdef H5F_RDCC_NEW
- if (idx_hint) *idx_hint = found;
-#endif
+ if (idx_hint) *idx_hint = idx;
ret_value = chunk;
done:
@@ -1370,25 +1415,14 @@ H5F_istore_unlock (H5F_t *f, const H5O_layout_t *layout,
FUNC_ENTER (H5F_istore_unlock, FAIL);
-#ifdef H5F_RDCC_NEW
if (-999==*idx_hint) {
/*not in cache*/
} else {
assert(*idx_hint>=0 && *idx_hint<rdcc->nslots);
- assert(rdcc->slot[*idx_hint].chunk==chunk);
+ assert(rdcc->slot[*idx_hint]);
+ assert(rdcc->slot[*idx_hint]->chunk==chunk);
found = *idx_hint;
}
-#else
- /* First look at the hint */
- if (idx_hint && *idx_hint>=0 && *idx_hint<rdcc->nused) {
- if (rdcc->slot[*idx_hint].chunk==chunk) found = *idx_hint;
- }
-
- /* Then look at all the entries */
- for (i=0; found<0 && i<rdcc->nused; i++) {
- if (rdcc->slot[i].chunk==chunk) found = i;
- }
-#endif
if (found<0) {
/*
@@ -1417,7 +1451,7 @@ H5F_istore_unlock (H5F_t *f, const H5O_layout_t *layout,
/*
* It's in the cache so unlock it.
*/
- ent = rdcc->slot + found;
+ ent = rdcc->slot[found];
assert (ent->locked);
if (dirty) {
ent->dirty = TRUE;
diff --git a/src/H5F.c b/src/H5F.c
index a3e6782..25c328c 100644
--- a/src/H5F.c
+++ b/src/H5F.c
@@ -159,6 +159,7 @@ H5F_init_interface(void)
/* Initialize the default file access property list */
H5F_access_dflt.mdc_nelmts = H5AC_NSLOTS;
+ H5F_access_dflt.rdcc_nelmts = 521;
H5F_access_dflt.rdcc_nbytes = 1024*1024; /*1MB*/
H5F_access_dflt.rdcc_w0 = 0.75; /*preempt fully read chunks*/
H5F_access_dflt.threshold = 1; /*alignment applies to everything*/
diff --git a/src/H5Fistore.c b/src/H5Fistore.c
index dcd1d35..278bca9 100644
--- a/src/H5Fistore.c
+++ b/src/H5Fistore.c
@@ -13,6 +13,21 @@
* chunk index to disk address. Each chunk can be compressed
* independently and the chunks may move around in the file as
* their storage requirements change.
+ *
+ * Cache: Disk I/O is performed in units of chunks and H5MF_alloc()
+ * contains code to optionally align chunks on disk block
+ * boundaries for performance.
+ *
+ * The chunk cache is an extendible hash indexed by a function
+ * of storage B-tree address and chunk N-dimensional offset
+ * within the dataset. Collisions are not resolved -- one of
+ * the two chunks competing for the hash slot must be preempted
+ * from the cache. All entries in the hash also participate in
+ * a doubly-linked list and entries are penalized by moving them
+ * toward the front of the list. When a new chunk is about to
+ * be added to the cache the heap is pruned by preempting
+ * entries near the front of the list to make room for the new
+ * entry which is added to the end of the list.
*/
#include <H5private.h>
#include <H5Dprivate.h>
@@ -23,6 +38,31 @@
#include <H5Oprivate.h>
#include <H5Vprivate.h>
+/*
+ * Feature: If this constant is defined then every cache preemption and load
+ * causes a character to be printed on the standard error stream:
+ *
+ * `.': Entry was preempted because it has been completely read or
+ * completely written but not partially read and not partially
+ * written. This is often a good reason for preemption because such
+ * a chunk will be unlikely to be referenced in the near future.
+ *
+ * `:': Entry was preempted because it hasn't been used recently.
+ *
+ * `#': Entry was preempted because another chunk collided with it. This
+ * is usually a relatively bad thing. If there are too many of
+ * these then the number of entries in the cache can be increased.
+ *
+ * c: Entry was preempted because the file is closing.
+ *
+ * w: A chunk read operation was eliminated because the library is
+ * about to write new values to the entire chunk. This is a good
+ * thing, especially on files where the chunk size is the same as
+ * the disk block size, chunks are aligned on disk block boundaries,
+ * and the operating system can also eliminate a read operation.
+ */
+/* #define H5F_ISTORE_DEBUG */
+
/* Interface initialization */
#define PABLO_MASK H5F_istore_mask
static hbool_t interface_initialize_g = FALSE;
@@ -40,6 +80,9 @@ typedef struct H5F_rdcc_ent_t {
size_t chunk_size; /*size of a chunk */
size_t alloc_size; /*amount allocated for the chunk */
uint8 *chunk; /*the unfiltered chunk data */
+ intn idx; /*index in hash table */
+ struct H5F_rdcc_ent_t *next;/*next item in doubly-linked list */
+ struct H5F_rdcc_ent_t *prev;/*previous item in doubly-linked list */
} H5F_rdcc_ent_t;
/* Private prototypes */
@@ -688,9 +731,10 @@ H5F_istore_init (H5F_t *f)
FUNC_ENTER (H5F_istore_init, FAIL);
HDmemset (rdcc, 0, sizeof(H5F_rdcc_t));
- if (f->shared->access_parms->rdcc_nbytes>0) {
- rdcc->nslots = 25; /*some initial number of slots*/
- rdcc->slot = H5MM_calloc (rdcc->nslots*sizeof(H5F_rdcc_ent_t));
+ if (f->shared->access_parms->rdcc_nbytes>0 &&
+ f->shared->access_parms->rdcc_nelmts>0) {
+ rdcc->nslots = f->shared->access_parms->rdcc_nelmts;
+ rdcc->slot = H5MM_calloc (rdcc->nslots*sizeof(H5F_rdcc_ent_t*));
if (NULL==rdcc->slot) {
HRETURN_ERROR (H5E_RESOURCE, H5E_NOSPACE, FAIL,
"memory allocation failed");
@@ -797,7 +841,7 @@ H5F_istore_flush_entry (H5F_t *f, H5F_rdcc_ent_t *ent, hbool_t reset)
f->shared->rdcc.nflushes++;
}
- /* Reset */
+ /* Reset, but do not free or removed from list */
if (reset) {
point_of_no_return = FALSE;
ent->layout = H5O_free(H5O_LAYOUT, ent->layout);
@@ -815,7 +859,8 @@ H5F_istore_flush_entry (H5F_t *f, H5F_rdcc_ent_t *ent, hbool_t reset)
/*
* If we reached the point of no return then we have no choice but to
* reset the entry. This can only happen if RESET is true but the
- * output pipeline failed.
+ * output pipeline failed. Do not free the entry or remove it from the
+ * list.
*/
if (ret_value<0 && point_of_no_return) {
ent->layout = H5O_free(H5O_LAYOUT, ent->layout);
@@ -846,25 +891,17 @@ H5F_istore_flush_entry (H5F_t *f, H5F_rdcc_ent_t *ent, hbool_t reset)
herr_t
H5F_istore_flush (H5F_t *f)
{
- H5F_rdcc_t *rdcc = &(f->shared->rdcc);
- intn i, nerrors=0;
+ H5F_rdcc_t *rdcc = &(f->shared->rdcc);
+ intn nerrors=0;
+ H5F_rdcc_ent_t *ent=NULL;
FUNC_ENTER (H5F_istore_flush, FAIL);
-#ifdef H5F_RDCC_NEW
- for (i=0; i<rdcc->nslots; i++) {
- if (rdcc->slot[i].chunk &&
- H5F_istore_flush_entry(f, rdcc->slot+i, FALSE)<0) {
- nerrors++;
- }
- }
-#else
- for (i=0; i<rdcc->nused; i++) {
- if (H5F_istore_flush_entry (f, rdcc->slot+i, FALSE)<0) {
+ for (ent=rdcc->head; ent; ent=ent->next) {
+ if (H5F_istore_flush_entry(f, ent, FALSE)<0) {
nerrors++;
}
}
-#endif
if (nerrors) {
HRETURN_ERROR (H5E_IO, H5E_CANTFLUSH, FAIL,
"unable to flush one or more raw data chunks");
@@ -890,29 +927,41 @@ H5F_istore_flush (H5F_t *f)
*-------------------------------------------------------------------------
*/
static herr_t
-H5F_istore_preempt (H5F_t *f, intn idx)
+H5F_istore_preempt (H5F_t *f, H5F_rdcc_ent_t *ent)
{
H5F_rdcc_t *rdcc = &(f->shared->rdcc);
- H5F_rdcc_ent_t *ent = rdcc->slot + idx;
FUNC_ENTER (H5F_istore_preempt, FAIL);
-#ifdef H5F_RDCC_NEW
- assert(idx>=0 && idx<rdcc->nslots);
- assert (!ent->locked);
+ assert(f);
+ assert(ent);
+ assert(!ent->locked);
+ assert(ent->idx>=0 && ent->idx<rdcc->nslots);
+ /* Flush */
H5F_istore_flush_entry(f, ent, TRUE);
- rdcc->nbytes -= ent->chunk_size;
-#else
- assert (idx>=0 && idx<rdcc->nused);
- assert (!ent->locked);
- H5F_istore_flush_entry (f, ent, TRUE);
- HDmemmove (rdcc->slot+idx, rdcc->slot+idx+1,
- (rdcc->nused-(idx+1)) * sizeof(H5F_rdcc_ent_t));
- rdcc->nused -= 1;
+ /* Unlink from list */
+ if (ent->prev) {
+ ent->prev->next = ent->next;
+ } else {
+ rdcc->head = ent->next;
+ }
+ if (ent->next) {
+ ent->next->prev = ent->prev;
+ } else {
+ rdcc->tail = ent->prev;
+ }
+ ent->prev = ent->next = NULL;
+
+ /* Remove from cache */
+ rdcc->slot[ent->idx] = NULL;
+ ent->idx = -1;
rdcc->nbytes -= ent->chunk_size;
-#endif
+ --rdcc->nused;
+
+ /* Free */
+ H5MM_xfree(ent);
FUNC_LEAVE (SUCCEED);
}
@@ -938,25 +987,22 @@ H5F_istore_preempt (H5F_t *f, intn idx)
herr_t
H5F_istore_dest (H5F_t *f)
{
- H5F_rdcc_t *rdcc = &(f->shared->rdcc);
- intn i, nerrors=0;
+ H5F_rdcc_t *rdcc = &(f->shared->rdcc);
+ intn nerrors=0;
+ H5F_rdcc_ent_t *ent=NULL, *next=NULL;
FUNC_ENTER (H5F_istore_dest, FAIL);
-#ifdef H5F_RDCC_NEW
- for (i=0; i<rdcc->nslots; i++) {
- if (rdcc->slot[i].chunk &&
- H5F_istore_preempt(f, i)<0) {
- nerrors++;
- }
- }
-#else
- for (i=rdcc->nused-1; i>=0; --i) {
- if (H5F_istore_preempt(f, i)<0) {
+ for (ent=rdcc->head; ent; ent=next) {
+#ifdef H5F_ISTORE_DEBUG
+ fputc('c', stderr);
+ fflush(stderr);
+#endif
+ next = ent->next;
+ if (H5F_istore_preempt(f, ent)<0) {
nerrors++;
}
}
-#endif
if (nerrors) {
HRETURN_ERROR (H5E_IO, H5E_CANTFLUSH, FAIL,
"unable to flush one or more raw data chunks");
@@ -989,71 +1035,89 @@ H5F_istore_dest (H5F_t *f)
static herr_t
H5F_istore_prune (H5F_t *f, size_t size)
{
-#ifdef H5F_RDCC_NEW
- intn i, nerrors=0;
- static intn place=0;
- H5F_rdcc_t *rdcc = &(f->shared->rdcc);
- H5F_rdcc_ent_t *ent = NULL;
- size_t total = f->shared->access_parms->rdcc_nbytes;
-#else
- intn i, meth0, meth1, nerrors=0;
+ intn i, j, nerrors=0;
H5F_rdcc_t *rdcc = &(f->shared->rdcc);
- H5F_rdcc_ent_t *ent0, *ent1;
- double w0 = f->shared->access_parms->rdcc_w0;
size_t total = f->shared->access_parms->rdcc_nbytes;
-#endif
-
+ const int nmeth=2; /*number of methods */
+ intn w[1]; /*weighting as an interval */
+ H5F_rdcc_ent_t *p[2], *cur; /*list pointers */
+ H5F_rdcc_ent_t *n[2]; /*list next pointers */
+
FUNC_ENTER (H5F_istore_prune, FAIL);
-#ifdef H5F_RDCC_NEW
- for (i=0; rdcc->nbytes+size>total && i<rdcc->nslots; i++, place++) {
- if (place>=rdcc->nslots) place = 0;
- ent = rdcc->slot+place;
- if (ent->chunk && !ent->locked) {
- if (H5F_istore_preempt(f, place)<0) nerrors++;
- }
- }
-#else
/*
- * We have two pointers that slide down the cache beginning at the least
- * recently used entry. The distance between the pointers represents the
- * relative weight. A weight of 50% for the first pointer means that the
- * second pointer is half the cache length behind the first pointer.
+ * Preemption is accomplished by having multiple pointers (currently two)
+ * slide down the list beginning at the head. Pointer p(N+1) will start
+ * traversing the list when pointer pN reaches wN percent of the original
+ * list. In other words, preemption method N gets to consider entries in
+ * approximate least recently used order w0 percent before method N+1
+ * where 100% means tha method N will run to completion before method N+1
+ * begins. The pointers participating in the list traversal are each
+ * given a chance at preemption before any of the pointers are advanced.
*/
- meth0 = rdcc->nused;
- meth1 = rdcc->nused * (1.0+w0);
- for (i=MAX(meth0, meth1)-1;
- rdcc->nbytes+size>total && i>=0;
- --i, --meth0, --meth1) {
-
- ent0 = rdcc->slot+meth0; /*might be a bad pointer!*/
- ent1 = rdcc->slot+meth1; /*might be a bad pointer!*/
+ w[0] = rdcc->nused * f->shared->access_parms->rdcc_w0;
+ p[0] = rdcc->head;
+ p[1] = NULL;
+
+ while ((p[0] || p[1]) && rdcc->nbytes+size>total) {
+
+ /* Introduce new pointers */
+ for (i=0; i<nmeth-1; i++) if (0==w[i]) p[i+1] = rdcc->head;
- if (meth0>=0 && meth0<rdcc->nused && !ent0->locked &&
- (0==ent0->rd_count || ent0->chunk_size==ent0->rd_count) &&
- (0==ent0->wr_count || ent0->chunk_size==ent0->wr_count)) {
- /*
- * Method 0: Preempt entries that have a zero rd_count. If the
- * application is accessing a dataset with a set of
- * non-overlapping partial I/O requests then chunks with a zero
- * rd_count will probably not be accessed in the near future.
- */
- if (H5F_istore_preempt (f, meth0)<0) nerrors++;
-
- } else if (meth1>=0 && meth1<rdcc->nused && !ent1->locked) {
- /*
- * Method 1: Discard the least recently used members from the
- * cache. This is a catch-all.
- */
- if (H5F_istore_preempt (f, meth1)<0) nerrors++;
+ /* Compute next value for each pointer */
+ for (i=0; i<nmeth; i++) n[i] = p[i] ? p[i]->next : NULL;
+
+ /* Give each method a chance */
+ for (i=0; i<nmeth && rdcc->nbytes+size>total; i++) {
+ if (0==i && p[0] && !p[0]->locked &&
+ ((0==p[0]->rd_count && 0==p[0]->wr_count) ||
+ (0==p[0]->rd_count && p[0]->chunk_size==p[0]->wr_count) ||
+ (p[0]->chunk_size==p[0]->rd_count && 0==p[0]->wr_count))) {
+ /*
+ * Method 0: Preempt entries that have been completely written
+ * and/or completely read but not entries that are partially
+ * written or partially read.
+ */
+ cur = p[0];
+#ifdef H5F_ISTORE_DEBUG
+ putc('.', stderr);
+ fflush(stderr);
+#endif
+
+ } else if (1==i && p[1] && !p[1]->locked) {
+ /*
+ * Method 1: Preempt the entry without regard to
+ * considerations other than being locked. This is the last
+ * resort preemption.
+ */
+ cur = p[1];
+#ifdef H5F_ISTORE_DEBUG
+ putc(':', stderr);
+ fflush(stderr);
+#endif
+
+ } else {
+ /* Nothing to preempt at this point */
+ cur= NULL;
+ }
+
+ if (cur) {
+ if (H5F_istore_preempt(f, cur)<0) nerrors++;
+ for (j=0; j<nmeth; j++) {
+ if (p[j]==cur) p[j] = NULL;
+ }
+ }
}
+
+ /* Advance pointers */
+ for (i=0; i<nmeth; i++) p[i] = n[i];
+ for (i=0; i<nmeth-1; i++) w[i] -= 1;
}
-#endif
+
if (nerrors) {
HRETURN_ERROR (H5E_IO, H5E_CANTFLUSH, FAIL,
"unable to preempt one or more raw data cache entry");
}
-
FUNC_LEAVE (SUCCEED);
}
@@ -1092,12 +1156,11 @@ H5F_istore_lock (H5F_t *f, const H5O_layout_t *layout,
const H5O_pline_t *pline, const hssize_t offset[],
hbool_t relax, intn *idx_hint/*in,out*/)
{
-#ifdef H5F_RDCC_NEW
- uintn idx;
-#endif
+ uintn idx; /*hash index number */
+ hbool_t found = FALSE; /*already in cache? */
H5F_rdcc_t *rdcc = &(f->shared->rdcc);/*raw data chunk cache*/
H5F_rdcc_ent_t *ent = NULL; /*cache entry */
- intn i, j, found = -1; /*counters */
+ intn i; /*counters */
H5F_istore_ud1_t udata; /*B-tree pass-through */
size_t chunk_size=0; /*size of a chunk */
size_t chunk_alloc=0; /*allocated chunk size */
@@ -1107,61 +1170,41 @@ H5F_istore_lock (H5F_t *f, const H5O_layout_t *layout,
FUNC_ENTER (H5F_istore_lock, NULL);
-#ifdef H5F_RDCC_NEW
if (rdcc->nslots>0) {
idx = layout->addr.offset;
- for (i=0; i<layout->ndims; i++) idx ^= offset[i];
+ for (i=0; i<layout->ndims; i++) idx = (idx<<1) ^ offset[i];
idx %= rdcc->nslots;
- ent = rdcc->slot + idx;
+ ent = rdcc->slot[idx];
- if (ent->chunk &&
+ if (ent &&
layout->ndims==ent->layout->ndims &&
H5F_addr_eq(&(layout->addr), &(ent->layout->addr))) {
- for (i=0, found=idx; i<ent->layout->ndims; i++) {
+ for (i=0, found=TRUE; i<ent->layout->ndims; i++) {
if (offset[i]!=ent->offset[i]) {
- found = -1;
+ found = FALSE;
break;
}
}
}
}
-#else
- /* First use the hint because that's O(1) */
- if (idx_hint && *idx_hint>=0 && *idx_hint<rdcc->nused) {
- ent = rdcc->slot + *idx_hint;
- if (layout->ndims==ent->layout->ndims &&
- H5F_addr_eq(&(layout->addr), &(ent->layout->addr))) {
- for (i=0, found=*idx_hint; found>=0 && i<ent->layout->ndims; i++) {
- if (offset[i]!=ent->offset[i]) found = -1;
- }
- }
- }
-
- /* If the hint is wrong then search the cache, O(n) */
- for (i=0; found<0 && i<rdcc->nused; i++) {
- ent = rdcc->slot + i;
- if (layout->ndims==ent->layout->ndims &&
- H5F_addr_eq(&(layout->addr), &(ent->layout->addr))) {
- for (j=0, found=i; found>=0 && j<ent->layout->ndims; j++) {
- if (offset[j]!=ent->offset[j]) found = -1;
- }
- }
- }
-#endif
- if (found>=0) {
+ if (found) {
/*
* Already in the cache. Count a hit.
*/
rdcc->nhits++;
-
- } else if (found<0 && relax) {
+
+ } else if (!found && relax) {
/*
* Not in the cache, but we're about to overwrite the whole thing
* anyway, so just allocate a buffer for it but don't initialize that
* buffer with the file contents. Count this as a hit instead of a
* miss because we saved ourselves lots of work.
*/
+#ifdef H5F_ISTORE_DEBUG
+ putc('w', stderr);
+ fflush(stderr);
+#endif
rdcc->nhits++;
for (i=0, chunk_size=1; i<layout->ndims; i++) {
chunk_size *= layout->dim[i];
@@ -1215,26 +1258,44 @@ H5F_istore_lock (H5F_t *f, const H5O_layout_t *layout,
rdcc->ninits++;
}
}
-
- assert (found>=0 || chunk_size>0);
-#ifdef H5F_RDCC_NEW
- if (found<0 && rdcc->nslots>0 &&
+ assert (found || chunk_size>0);
+
+ if (!found && rdcc->nslots>0 &&
chunk_size<=f->shared->access_parms->rdcc_nbytes &&
- (NULL==ent->chunk || !ent->locked)) {
+ (!ent || !ent->locked)) {
/*
* Add the chunk to the cache only if the slot is not already locked.
* Preempt enough things from the cache to make room.
*/
+ if (ent) {
+#ifdef H5F_ISTORE_DEBUG
+ putc('#', stderr);
+ fflush(stderr);
+#endif
+#if 0
+ HDfprintf(stderr, "\ncollision %3d %10a {",
+ idx, &(ent->layout->addr));
+ for (i=0; i<layout->ndims; i++) {
+ HDfprintf(stderr, "%s%Zu", i?",":"", ent->offset[i]);
+ }
+ HDfprintf(stderr, "}\n %10a {", &(layout->addr));
+ for (i=0; i<layout->ndims; i++) {
+ HDfprintf(stderr, "%s%Zu", i?",":"", offset[i]);
+ }
+ fprintf(stderr, "}\n");
+#endif
+ if (H5F_istore_preempt(f, ent)<0) {
+ HGOTO_ERROR(H5E_IO, H5E_CANTINIT, NULL,
+ "unable to preempt chunk from cache");
+ }
+ }
if (H5F_istore_prune(f, chunk_size)<0) {
HGOTO_ERROR(H5E_IO, H5E_CANTINIT, NULL,
"unable to preempt chunk(s) from cache");
}
- if (rdcc->slot[idx].chunk &&
- H5F_istore_preempt(f, idx)<0) {
- HGOTO_ERROR(H5E_IO, H5E_CANTINIT, NULL,
- "unable to preempt chunk from cache");
- }
- ent = rdcc->slot + idx;
+
+ /* Create a new entry */
+ ent = H5MM_malloc(sizeof(H5F_rdcc_ent_t));
ent->locked = 0;
ent->dirty = FALSE;
ent->chunk_size = chunk_size;
@@ -1247,83 +1308,67 @@ H5F_istore_lock (H5F_t *f, const H5O_layout_t *layout,
ent->rd_count = chunk_size;
ent->wr_count = chunk_size;
ent->chunk = chunk;
- found = idx;
- }
-#else
- if (found<0 && chunk_size<=f->shared->access_parms->rdcc_nbytes) {
- /*
- * Add the chunk to the beginning of the cache after pruning the cache
- * to make room.
- */
- if (H5F_istore_prune (f, chunk_size)<0) {
- H5E_clear ();
- }
- if (rdcc->nused>=rdcc->nslots) {
- size_t na = MAX (25, 2*rdcc->nslots);
- H5F_rdcc_ent_t *x = H5MM_realloc (rdcc->slot,
- na*sizeof(H5F_rdcc_ent_t));
- if (NULL==x) {
- HGOTO_ERROR (H5E_RESOURCE, H5E_NOSPACE, NULL,
- "memory allocation failed");
- }
- rdcc->nslots = (intn)na;
- rdcc->slot = x;
- }
- HDmemmove (rdcc->slot+1, rdcc->slot,
- rdcc->nused*sizeof(H5F_rdcc_ent_t));
- rdcc->nused++;
+
+ /* Add it to the cache */
+ assert(NULL==rdcc->slot[idx]);
+ rdcc->slot[idx] = ent;
+ ent->idx = idx;
rdcc->nbytes += chunk_size;
- ent = rdcc->slot;
- ent->locked = 0;
- ent->dirty = FALSE;
- ent->chunk_size = chunk_size;
- ent->alloc_size = chunk_alloc;
- ent->layout = H5O_copy (H5O_LAYOUT, layout, NULL);
- ent->pline = H5O_copy (H5O_PLINE, pline, NULL);
- for (i=0; i<layout->ndims; i++) {
- ent->offset[i] = offset[i];
+ rdcc->nused++;
+
+ /* Add it to the linked list */
+ ent->next = NULL;
+ if (rdcc->tail) {
+ rdcc->tail->next = ent;
+ ent->prev = rdcc->tail;
+ rdcc->tail = ent;
+ } else {
+ rdcc->head = rdcc->tail = ent;
+ ent->prev = NULL;
}
- ent->rd_count = chunk_size;
- ent->wr_count = chunk_size;
- ent->chunk = chunk;
- found = 0;
- }
-#endif
- else if (found<0) {
+ found = TRUE;
+
+ } else if (!found) {
/*
* The chunk is larger than the entire cache so we don't cache it.
* This is the reason all those arguments have to be repeated for the
* unlock function.
*/
ent = NULL;
- found = -999;
+ idx = -999;
-#ifndef H5F_RDCC_NEW
- } else if (found>0) {
+ } else if (found) {
/*
- * The chunk is not at the beginning of the cache; move it forward by
- * one slot. This is how we implement the LRU preemption algorithm.
+ * The chunk is not at the beginning of the cache; move it backward
+ * by one slot. This is how we implement the LRU preemption
+ * algorithm.
*/
- H5F_rdcc_ent_t x = rdcc->slot[found];
- rdcc->slot[found] = rdcc->slot[found-1];
- rdcc->slot[found-1] = x;
- ent = rdcc->slot + --found;
-#endif
+ if (ent->next) {
+ if (ent->next->next) {
+ ent->next->next->prev = ent;
+ } else {
+ rdcc->tail = ent;
+ }
+ ent->next->prev = ent->prev;
+ if (ent->prev) {
+ ent->prev->next = ent->next;
+ } else {
+ rdcc->head = ent->next;
+ }
+ ent->prev = ent->next;
+ ent->next = ent->next->next;
+ ent->prev->next = ent;
+ }
}
/* Lock the chunk into the cache */
if (ent) {
assert (!ent->locked);
ent->locked = TRUE;
-#ifndef H5F_RDCC_NEW
- if (idx_hint) *idx_hint = found;
-#endif
chunk = ent->chunk;
}
-#ifdef H5F_RDCC_NEW
- if (idx_hint) *idx_hint = found;
-#endif
+ if (idx_hint) *idx_hint = idx;
ret_value = chunk;
done:
@@ -1370,25 +1415,14 @@ H5F_istore_unlock (H5F_t *f, const H5O_layout_t *layout,
FUNC_ENTER (H5F_istore_unlock, FAIL);
-#ifdef H5F_RDCC_NEW
if (-999==*idx_hint) {
/*not in cache*/
} else {
assert(*idx_hint>=0 && *idx_hint<rdcc->nslots);
- assert(rdcc->slot[*idx_hint].chunk==chunk);
+ assert(rdcc->slot[*idx_hint]);
+ assert(rdcc->slot[*idx_hint]->chunk==chunk);
found = *idx_hint;
}
-#else
- /* First look at the hint */
- if (idx_hint && *idx_hint>=0 && *idx_hint<rdcc->nused) {
- if (rdcc->slot[*idx_hint].chunk==chunk) found = *idx_hint;
- }
-
- /* Then look at all the entries */
- for (i=0; found<0 && i<rdcc->nused; i++) {
- if (rdcc->slot[i].chunk==chunk) found = i;
- }
-#endif
if (found<0) {
/*
@@ -1417,7 +1451,7 @@ H5F_istore_unlock (H5F_t *f, const H5O_layout_t *layout,
/*
* It's in the cache so unlock it.
*/
- ent = rdcc->slot + found;
+ ent = rdcc->slot[found];
assert (ent->locked);
if (dirty) {
ent->dirty = TRUE;
diff --git a/src/H5Fprivate.h b/src/H5Fprivate.h
index cf92283..6e80fa0 100644
--- a/src/H5Fprivate.h
+++ b/src/H5Fprivate.h
@@ -239,7 +239,8 @@ typedef struct H5F_create_t {
* File-access property list.
*/
typedef struct H5F_access_t {
- intn mdc_nelmts; /* Size of meta data cache (nelmts) */
+ intn mdc_nelmts; /* Size of meta data cache (elements) */
+ intn rdcc_nelmts; /* Size of raw data chunk cache (elmts) */
size_t rdcc_nbytes; /* Size of raw data chunk cache (bytes) */
double rdcc_w0; /* Preempt read chunks first? [0.0..1.0]*/
hsize_t threshold; /* Threshold for alignment */
@@ -419,10 +420,10 @@ typedef struct H5F_rdcc_t {
uintn nflushes;/* Number of cache flushes */
size_t nbytes; /* Current cached raw data in bytes */
intn nslots; /* Number of chunk slots allocated */
-#ifndef H5F_RDCC_NEW
+ struct H5F_rdcc_ent_t *head; /* Head of doubly linked list */
+ struct H5F_rdcc_ent_t *tail; /* Tail of doubly linked list */
intn nused; /* Number of chunk slots in use */
-#endif
- struct H5F_rdcc_ent_t *slot; /* Chunk slots, each points to a chunk */
+ struct H5F_rdcc_ent_t **slot; /* Chunk slots, each points to a chunk*/
} H5F_rdcc_t;
/*
diff --git a/src/H5P.c b/src/H5P.c
index 99b4e67..c229e0a 100644
--- a/src/H5P.c
+++ b/src/H5P.c
@@ -1878,14 +1878,16 @@ H5Pget_family(hid_t plist_id, hsize_t *memb_size/*out*/,
* Function: H5Pset_cache
*
* Purpose: Set the number of objects in the meta data cache and the
- * total number of bytes in the raw data chunk cache.
+ * maximum number of chunks and bytes in the raw data chunk
+ * cache.
*
* The RDCC_W0 value should be between 0 and 1 inclusive and
- * indicates how much chunks that have been fully read are
- * favored for preemption. A value of zero means fully read
- * chunks are treated no differently than other chunks (the
- * preemption is strictly LRU) while a value of one means fully
- * read chunks are always preempted before other chunks.
+ * indicates how much chunks that have been fully read or fully
+ * written are favored for preemption. A value of zero means
+ * fully read or written chunks are treated no differently than
+ * other chunks (the preemption is strictly LRU) while a value
+ * of one means fully read chunks are always preempted before
+ * other chunks.
*
* Return: Success: SUCCEED
*
@@ -1899,13 +1901,14 @@ H5Pget_family(hid_t plist_id, hsize_t *memb_size/*out*/,
*-------------------------------------------------------------------------
*/
herr_t
-H5Pset_cache(hid_t plist_id, int mdc_nelmts, size_t rdcc_nbytes,
- double rdcc_w0)
+H5Pset_cache(hid_t plist_id, int mdc_nelmts,
+ int rdcc_nelmts, size_t rdcc_nbytes, double rdcc_w0)
{
H5F_access_t *fapl = NULL;
FUNC_ENTER (H5Pset_cache, FAIL);
- H5TRACE4("e","iIszd",plist_id,mdc_nelmts,rdcc_nbytes,rdcc_w0);
+ H5TRACE5("e","iIsIszd",plist_id,mdc_nelmts,rdcc_nelmts,rdcc_nbytes,
+ rdcc_w0);
/* Check arguments */
if (H5P_FILE_ACCESS!=H5P_get_class (plist_id) ||
@@ -1917,6 +1920,10 @@ H5Pset_cache(hid_t plist_id, int mdc_nelmts, size_t rdcc_nbytes,
HRETURN_ERROR (H5E_ARGS, H5E_BADVALUE, FAIL,
"meta data cache size must be non-negative");
}
+ if (rdcc_nelmts<0) {
+ HRETURN_ERROR(H5E_ARGS, H5E_BADVALUE, FAIL,
+ "raw data chunk cache nelmts must be non-negative");
+ }
if (rdcc_w0<0.0 || rdcc_w0>1.0) {
HRETURN_ERROR (H5E_ARGS, H5E_BADVALUE, FAIL,
"raw data cache w0 value must be between 0.0 and 1.0 "
@@ -1925,6 +1932,7 @@ H5Pset_cache(hid_t plist_id, int mdc_nelmts, size_t rdcc_nbytes,
/* Set sizes */
fapl->mdc_nelmts = mdc_nelmts;
+ fapl->rdcc_nelmts = rdcc_nelmts;
fapl->rdcc_nbytes = rdcc_nbytes;
fapl->rdcc_w0 = rdcc_w0;
@@ -1936,10 +1944,10 @@ H5Pset_cache(hid_t plist_id, int mdc_nelmts, size_t rdcc_nbytes,
* Function: H5Pget_cache
*
* Purpose: Retrieves the maximum possible number of elements in the meta
- * data cache and the maximum possible number of bytes and the
- * RDCC_W0 value in the raw data chunk cache. Any (or all)
- * arguments may be null pointers in which case the corresponding
- * datum is not returned.
+ * data cache and the maximum possible number of elements and
+ * bytes and the RDCC_W0 value in the raw data chunk cache. Any
+ * (or all) arguments may be null pointers in which case the
+ * corresponding datum is not returned.
*
* Return: Success: SUCCEED
*
@@ -1953,13 +1961,14 @@ H5Pset_cache(hid_t plist_id, int mdc_nelmts, size_t rdcc_nbytes,
*-------------------------------------------------------------------------
*/
herr_t
-H5Pget_cache(hid_t plist_id, int *mdc_nelmts, size_t *rdcc_nbytes,
- double *rdcc_w0)
+H5Pget_cache(hid_t plist_id, int *mdc_nelmts,
+ int *rdcc_nelmts, size_t *rdcc_nbytes, double *rdcc_w0)
{
H5F_access_t *fapl = NULL;
FUNC_ENTER (H5Pget_cache, FAIL);
- H5TRACE4("e","i*Is*z*d",plist_id,mdc_nelmts,rdcc_nbytes,rdcc_w0);
+ H5TRACE5("e","i*Is*Is*z*d",plist_id,mdc_nelmts,rdcc_nelmts,rdcc_nbytes,
+ rdcc_w0);
/* Check arguments */
if (H5P_FILE_ACCESS!=H5P_get_class (plist_id) ||
@@ -1970,6 +1979,7 @@ H5Pget_cache(hid_t plist_id, int *mdc_nelmts, size_t *rdcc_nbytes,
/* Get sizes */
if (mdc_nelmts) *mdc_nelmts = fapl->mdc_nelmts;
+ if (rdcc_nelmts) *rdcc_nelmts = fapl->rdcc_nelmts;
if (rdcc_nbytes) *rdcc_nbytes = fapl->rdcc_nbytes;
if (rdcc_w0) *rdcc_w0 = fapl->rdcc_w0;
@@ -2103,6 +2113,7 @@ H5Pset_hyper_cache(hid_t plist_id, unsigned cache, unsigned limit)
H5D_xfer_t *plist = NULL;
FUNC_ENTER (H5Pset_hyper_cache, FAIL);
+ H5TRACE3("e","iIuIu",plist_id,cache,limit);
/* Check arguments */
if (H5P_DATASET_XFER != H5P_get_class (plist_id) ||
@@ -2141,6 +2152,7 @@ H5Pget_hyper_cache(hid_t plist_id, unsigned *cache/*out*/, unsigned *limit/*out*
H5D_xfer_t *plist = NULL;
FUNC_ENTER (H5Pget_hyper_cache, 0);
+ H5TRACE3("e","ixx",plist_id,cache,limit);
/* Check arguments */
if (H5P_DATASET_XFER != H5P_get_class (plist_id) ||
diff --git a/src/H5Ppublic.h b/src/H5Ppublic.h
index d82bb83..123ec41 100644
--- a/src/H5Ppublic.h
+++ b/src/H5Ppublic.h
@@ -101,9 +101,10 @@ H5Z_filter_t H5Pget_filter(hid_t plist_id, int filter,
unsigned cd_values[]/*out*/,
size_t namelen, char name[]);
herr_t H5Pset_deflate (hid_t plist_id, unsigned aggression);
-herr_t H5Pset_cache (hid_t plist_id, int mdc_nelmts, size_t rdcc_nbytes,
- double rdcc_w0);
-herr_t H5Pget_cache (hid_t plist_id, int *mdc_nelmts, size_t *rdcc_nbytes,
+herr_t H5Pset_cache (hid_t plist_id, int mdc_nelmts, int rdcc_nelmts,
+ size_t rdcc_nbytes, double rdcc_w0);
+herr_t H5Pget_cache (hid_t plist_id, int *mdc_nelmts/*out*/,
+ int *rdcc_nelmts/*out*/, size_t *rdcc_nbytes/*out*/,
double *rdcc_w0);
herr_t H5Pset_hyper_cache(hid_t plist_id, unsigned cache, unsigned limit);
herr_t H5Pget_hyper_cache(hid_t plist_id, unsigned *cache, unsigned *limit);