parallel_step_size.patch
application/octet-stream
Filename: parallel_step_size.patch
Type: application/octet-stream
Part: 1
Patch
Format: unified
| File | + | − |
|---|---|---|
| src/backend/access/heap/heapam.c | 16 | 6 |
| src/backend/access/table/tableam.c | 32 | 3 |
| src/include/access/relscan.h | 12 | 1 |
| src/include/access/tableam.h | 2 | 0 |
diff --git a/src/backend/access/heap/heapam.c b/src/backend/access/heap/heapam.c
index 94eb37d48d..20a823bbad 100644
--- a/src/backend/access/heap/heapam.c
+++ b/src/backend/access/heap/heapam.c
@@ -520,12 +520,14 @@ heapgettup(HeapScanDesc scan,
{
ParallelBlockTableScanDesc pbscan =
(ParallelBlockTableScanDesc) scan->rs_base.rs_parallel;
+ ParallelBlockTableScanWork pbscanwork =
+ (ParallelBlockTableScanWork) scan->rs_base.rs_parallel_work;
table_block_parallelscan_startblock_init(scan->rs_base.rs_rd,
- pbscan);
+ pbscanwork, pbscan);
page = table_block_parallelscan_nextpage(scan->rs_base.rs_rd,
- pbscan);
+ pbscanwork, pbscan);
/* Other processes might have already finished the scan. */
if (page == InvalidBlockNumber)
@@ -720,9 +722,11 @@ heapgettup(HeapScanDesc scan,
{
ParallelBlockTableScanDesc pbscan =
(ParallelBlockTableScanDesc) scan->rs_base.rs_parallel;
+ ParallelBlockTableScanWork pbscanwork =
+ (ParallelBlockTableScanWork) scan->rs_base.rs_parallel_work;
page = table_block_parallelscan_nextpage(scan->rs_base.rs_rd,
- pbscan);
+ pbscanwork, pbscan);
finished = (page == InvalidBlockNumber);
}
else
@@ -834,12 +838,14 @@ heapgettup_pagemode(HeapScanDesc scan,
{
ParallelBlockTableScanDesc pbscan =
(ParallelBlockTableScanDesc) scan->rs_base.rs_parallel;
+ ParallelBlockTableScanWork pbscanwork =
+ (ParallelBlockTableScanWork) scan->rs_base.rs_parallel_work;
table_block_parallelscan_startblock_init(scan->rs_base.rs_rd,
- pbscan);
+ pbscanwork, pbscan);
page = table_block_parallelscan_nextpage(scan->rs_base.rs_rd,
- pbscan);
+ pbscanwork, pbscan);
/* Other processes might have already finished the scan. */
if (page == InvalidBlockNumber)
@@ -1019,9 +1025,11 @@ heapgettup_pagemode(HeapScanDesc scan,
{
ParallelBlockTableScanDesc pbscan =
(ParallelBlockTableScanDesc) scan->rs_base.rs_parallel;
+ ParallelBlockTableScanWork pbscanwork =
+ (ParallelBlockTableScanWork) scan->rs_base.rs_parallel_work;
page = table_block_parallelscan_nextpage(scan->rs_base.rs_rd,
- pbscan);
+ pbscanwork, pbscan);
finished = (page == InvalidBlockNumber);
}
else
@@ -1155,6 +1163,8 @@ heap_beginscan(Relation relation, Snapshot snapshot,
scan->rs_base.rs_nkeys = nkeys;
scan->rs_base.rs_flags = flags;
scan->rs_base.rs_parallel = parallel_scan;
+ scan->rs_base.rs_parallel_work =
+ palloc0(sizeof(ParallelBlockTableScanWorkData));
scan->rs_strategy = NULL; /* set in initscan */
/*
diff --git a/src/backend/access/table/tableam.c b/src/backend/access/table/tableam.c
index c814733b22..be84ac4947 100644
--- a/src/backend/access/table/tableam.c
+++ b/src/backend/access/table/tableam.c
@@ -404,10 +404,16 @@ table_block_parallelscan_reinitialize(Relation rel, ParallelTableScanDesc pscan)
* to set the startblock once.
*/
void
-table_block_parallelscan_startblock_init(Relation rel, ParallelBlockTableScanDesc pbscan)
+table_block_parallelscan_startblock_init(Relation rel,
+ ParallelBlockTableScanWork workspace,
+ ParallelBlockTableScanDesc pbscan)
{
BlockNumber sync_startpage = InvalidBlockNumber;
+ /* Reset the state we use for controlling allocation size. */
+ memset(workspace, 0, sizeof(*workspace));
+ workspace->phsw_step_size = 1;
+
retry:
/* Grab the spinlock. */
SpinLockAcquire(&pbscan->phs_mutex);
@@ -447,7 +453,9 @@ retry:
* backend gets an InvalidBlockNumber return.
*/
BlockNumber
-table_block_parallelscan_nextpage(Relation rel, ParallelBlockTableScanDesc pbscan)
+table_block_parallelscan_nextpage(Relation rel,
+ ParallelBlockTableScanWork workspace,
+ ParallelBlockTableScanDesc pbscan)
{
BlockNumber page;
uint64 nallocated;
@@ -467,7 +475,28 @@ table_block_parallelscan_nextpage(Relation rel, ParallelBlockTableScanDesc pbsca
* The actual page to return is calculated by adding the counter to the
* starting block number, modulo nblocks.
*/
- nallocated = pg_atomic_fetch_add_u64(&pbscan->phs_nallocated, 1);
+ if (workspace->phsw_nallocated_range > 0)
+ {
+ nallocated = ++workspace->phsw_nallocated;
+ workspace->phsw_nallocated_range--;
+ }
+ else
+ {
+ /*
+ * Ramp up the step size as we go, until we approach the end of the
+ * scan, and then ramp it back down again to avoid unfair allocation.
+ * XXX Come up with a better algorithm!
+ */
+ if (workspace->phsw_nallocated > pbscan->phs_nblocks - 64)
+ workspace->phsw_step_size = 1;
+ else if (workspace->phsw_step_size < 64)
+ workspace->phsw_step_size *= 2;
+
+ nallocated = workspace->phsw_nallocated =
+ pg_atomic_fetch_add_u64(&pbscan->phs_nallocated,
+ workspace->phsw_step_size);
+ workspace->phsw_nallocated_range = workspace->phsw_step_size - 1;
+ }
if (nallocated >= pbscan->phs_nblocks)
page = InvalidBlockNumber; /* all blocks have been allocated */
else
diff --git a/src/include/access/relscan.h b/src/include/access/relscan.h
index 6f0258831f..0d440c9de3 100644
--- a/src/include/access/relscan.h
+++ b/src/include/access/relscan.h
@@ -42,9 +42,9 @@ typedef struct TableScanDescData
*/
uint32 rs_flags;
+ void *rs_parallel_work;
struct ParallelTableScanDescData *rs_parallel; /* parallel scan
* information */
-
} TableScanDescData;
typedef struct TableScanDescData *TableScanDesc;
@@ -81,6 +81,17 @@ typedef struct ParallelBlockTableScanDescData
} ParallelBlockTableScanDescData;
typedef struct ParallelBlockTableScanDescData *ParallelBlockTableScanDesc;
+/*
+ * Per backend state for parallel table sacan, for block oriented storage.
+ */
+typedef struct ParallelBlockTableScanWorkData
+{
+ uint64 phsw_nallocated;
+ int phsw_nallocated_range;
+ int phsw_step_size;
+} ParallelBlockTableScanWorkData;
+typedef struct ParallelBlockTableScanWorkData *ParallelBlockTableScanWork;
+
/*
* Base class for fetches from a table via an index. This is the base-class
* for such scans, which needs to be embedded in the respective struct for
diff --git a/src/include/access/tableam.h b/src/include/access/tableam.h
index eb18739c36..f0cefd7e26 100644
--- a/src/include/access/tableam.h
+++ b/src/include/access/tableam.h
@@ -1790,8 +1790,10 @@ extern Size table_block_parallelscan_initialize(Relation rel,
extern void table_block_parallelscan_reinitialize(Relation rel,
ParallelTableScanDesc pscan);
extern BlockNumber table_block_parallelscan_nextpage(Relation rel,
+ ParallelBlockTableScanWork pbscanwork,
ParallelBlockTableScanDesc pbscan);
extern void table_block_parallelscan_startblock_init(Relation rel,
+ ParallelBlockTableScanWork pbscanwork,
ParallelBlockTableScanDesc pbscan);