Skip to content

Commit

Permalink
Refactor diskquota local_table_stats_map (#34)
Browse files Browse the repository at this point in the history 
During initialization, diskquota used a non-optimal structure for the local
hashmap local_table_stats_map. In a hashmap, there is quite a significant
overhead for each entry. Therefore, a large number of small entries led to
increased RAM consumption during cluster startup. Change the specified
structure, making the table oid as the key, and an array of sizes by segments
as the value. This significantly reduces the amount of memory consumed, because
now there will be SEGCOUNT times fewer records. Also fix a small bug with
duplicate oid tables in the active_oids string array in the dispatch_rejectmap
function.

Ticket: ADBDEV-6442
  • Loading branch information
@RekGRpth
RekGRpth authored Oct 17, 2024
1 parent 2862042 commit 24546b2
Show file tree
Hide file tree
Showing 3 changed files with 65 additions and 83 deletions.
83 changes: 28 additions & 55 deletions src/gp_activetable.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -374,12 +374,12 @@ gp_fetch_active_tables(bool is_init)
Assert(Gp_role == GP_ROLE_DISPATCH);

memset(&ctl, 0, sizeof(ctl));
ctl.keysize = sizeof(TableEntryKey);
ctl.entrysize = sizeof(DiskQuotaActiveTableEntry);
ctl.keysize = sizeof(Oid);
ctl.entrysize = sizeof(ActiveTableEntryCombined) + SEGCOUNT * sizeof(Size);
ctl.hcxt = CurrentMemoryContext;

local_table_stats_map = diskquota_hash_create("local active table map with relfilenode info", 1024, &ctl,
HASH_ELEM | HASH_CONTEXT, DISKQUOTA_TAG_HASH);
HASH_ELEM | HASH_CONTEXT, DISKQUOTA_OID_HASH);

if (is_init)
{
Expand Down Expand Up @@ -946,14 +946,13 @@ get_active_tables_oid(void)
static void
load_table_size(HTAB *local_table_stats_map)
{
TupleDesc tupdesc;
int i;
bool found;
TableEntryKey key;
DiskQuotaActiveTableEntry *quota_entry;
SPIPlanPtr plan;
Portal portal;
char *sql = "select tableid, size, segid from diskquota.table_size";
TupleDesc tupdesc;
int i;
bool found;
ActiveTableEntryCombined *quota_entry;
SPIPlanPtr plan;
Portal portal;
char *sql = "select tableid, size, segid from diskquota.table_size";

if ((plan = SPI_prepare(sql, 0, NULL)) == NULL)
ereport(ERROR, (errmsg("[diskquota] SPI_prepare(\"%s\") failed", sql)));
Expand Down Expand Up @@ -1016,14 +1015,11 @@ load_table_size(HTAB *local_table_stats_map)
size = DatumGetInt64(dat);
dat = SPI_getbinval(tup, tupdesc, 3, &isnull);
if (isnull) continue;
segid = DatumGetInt16(dat);
key.reloid = reloid;
key.segid = segid;

quota_entry = (DiskQuotaActiveTableEntry *)hash_search(local_table_stats_map, &key, HASH_ENTER, &found);
quota_entry->reloid = reloid;
quota_entry->tablesize = size;
quota_entry->segid = segid;
segid = DatumGetInt16(dat);

quota_entry = (ActiveTableEntryCombined *)hash_search(local_table_stats_map, &reloid, HASH_ENTER, &found);
quota_entry->reloid = reloid;
quota_entry->tablesize[segid + 1] = size;
}
SPI_freetuptable(SPI_tuptable);
SPI_cursor_fetch(portal, true, 10000);
Expand Down Expand Up @@ -1164,12 +1160,11 @@ pull_active_table_size_from_seg(HTAB *local_table_stats_map, char *active_oid_ar
/* sum table size from each segment into local_table_stats_map */
for (i = 0; i < cdb_pgresults.numResults; i++)
{
Size tableSize;
bool found;
Oid reloid;
int segId;
TableEntryKey key;
DiskQuotaActiveTableEntry *entry;
Size tableSize;
bool found;
Oid reloid;
int segId;
ActiveTableEntryCombined *entry;

PGresult *pgresult = cdb_pgresults.pg_results[i];

Expand All @@ -1182,42 +1177,20 @@ pull_active_table_size_from_seg(HTAB *local_table_stats_map, char *active_oid_ar

for (j = 0; j < PQntuples(pgresult); j++)
{
reloid = atooid(PQgetvalue(pgresult, j, 0));
tableSize = (Size)atoll(PQgetvalue(pgresult, j, 1));
key.reloid = reloid;
reloid = atooid(PQgetvalue(pgresult, j, 0));
tableSize = (Size)atoll(PQgetvalue(pgresult, j, 1));
entry = (ActiveTableEntryCombined *)hash_search(local_table_stats_map, &reloid, HASH_ENTER, &found);

/* for diskquota extension version is 1.0, pgresult doesn't contain segid */
if (PQnfields(pgresult) == 3)
{
/* get the segid, tablesize for each table */
segId = atoi(PQgetvalue(pgresult, j, 2));
key.segid = segId;
entry = (DiskQuotaActiveTableEntry *)hash_search(local_table_stats_map, &key, HASH_ENTER, &found);

if (!found)
{
/* receive table size info from the first segment */
entry->reloid = reloid;
entry->segid = segId;
}
entry->tablesize = tableSize;
segId = atoi(PQgetvalue(pgresult, j, 2));
entry->tablesize[segId + 1] = tableSize;
}

/* when segid is -1, the tablesize is the sum of tablesize of master and all segments */
key.segid = -1;
entry = (DiskQuotaActiveTableEntry *)hash_search(local_table_stats_map, &key, HASH_ENTER, &found);

if (!found)
{
/* receive table size info from the first segment */
entry->reloid = reloid;
entry->tablesize = tableSize;
entry->segid = -1;
}
else
{
/* sum table size from all the segments */
entry->tablesize = entry->tablesize + tableSize;
}
/* tablesize for index 0 is the sum of tablesize of master and all segments */
entry->tablesize[0] = (found ? entry->tablesize[0] : 0) + tableSize;
}
}
cdbdisp_clearCdbPgResults(&cdb_pgresults);
Expand Down
10 changes: 10 additions & 0 deletions src/gp_activetable.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -37,6 +37,16 @@ typedef struct DiskQuotaActiveTableEntry
Size tablesize;
} DiskQuotaActiveTableEntry;

typedef struct ActiveTableEntryCombined
{
Oid reloid;
/*
Variable length array: index 0 is used for the coordinator,
the remaining SEGCOUNT indexes are for segments.
*/
Size tablesize[1];
} ActiveTableEntryCombined;

extern HTAB *gp_fetch_active_tables(bool force);
extern void init_active_table_hook(void);
extern void init_shm_worker_active_tables(void);
Expand Down
55 changes: 27 additions & 28 deletions src/quotamodel.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -912,19 +912,18 @@ merge_uncommitted_table_to_oidlist(List *oidlist)
static void
calculate_table_disk_usage(bool is_init, HTAB *local_active_table_stat_map)
{
bool table_size_map_found;
bool active_tbl_found;
int64 updated_total_size;
TableSizeEntry *tsentry = NULL;
Oid relOid;
HASH_SEQ_STATUS iter;
DiskQuotaActiveTableEntry *active_table_entry;
TableSizeEntryKey key;
TableEntryKey active_table_key;
List *oidlist;
ListCell *l;
int delete_entries_num = 0;
StringInfoData delete_statement;
bool table_size_map_found;
bool active_tbl_found;
int64 updated_total_size;
TableSizeEntry *tsentry = NULL;
Oid relOid;
HASH_SEQ_STATUS iter;
ActiveTableEntryCombined *active_table_entry;
TableSizeEntryKey key;
List *oidlist;
ListCell *l;
int delete_entries_num = 0;
StringInfoData delete_statement;

initStringInfo(&delete_statement);

Expand Down Expand Up @@ -1042,10 +1041,8 @@ calculate_table_disk_usage(bool is_init, HTAB *local_active_table_stat_map)

/* mark tsentry is_exist */
if (tsentry) set_table_size_entry_flag(tsentry, TABLE_EXIST);
active_table_key.reloid = relOid;
active_table_key.segid = cur_segid;
active_table_entry = (DiskQuotaActiveTableEntry *)hash_search(
local_active_table_stat_map, &active_table_key, HASH_FIND, &active_tbl_found);
active_table_entry = (ActiveTableEntryCombined *)hash_search(local_active_table_stat_map, &relOid,
HASH_FIND, &active_tbl_found);

/* skip to recalculate the tables which are not in active list */
if (active_tbl_found)
Expand All @@ -1055,15 +1052,17 @@ calculate_table_disk_usage(bool is_init, HTAB *local_active_table_stat_map)
/* pretend process as utility mode, and append the table size on master */
Gp_role = GP_ROLE_UTILITY;

active_table_entry->tablesize += calculate_table_size(relOid);
/* when cur_segid is -1, the tablesize is the sum of tablesize of master and all segments */
active_table_entry->tablesize[0] += calculate_table_size(relOid);

Gp_role = GP_ROLE_DISPATCH;
}
/* firstly calculate the updated total size of a table */
updated_total_size = active_table_entry->tablesize - TableSizeEntryGetSize(tsentry, cur_segid);
updated_total_size =
active_table_entry->tablesize[cur_segid + 1] - TableSizeEntryGetSize(tsentry, cur_segid);

/* update the table_size entry */
TableSizeEntrySetSize(tsentry, cur_segid, active_table_entry->tablesize);
TableSizeEntrySetSize(tsentry, cur_segid, active_table_entry->tablesize[cur_segid + 1]);
TableSizeEntrySetFlushFlag(tsentry, cur_segid);

/* update the disk usage, there may be entries in the map whose keys are InvlidOid as the tsentry does
Expand Down Expand Up @@ -1345,14 +1344,14 @@ flush_local_reject_map(void)
static void
dispatch_rejectmap(HTAB *local_active_table_stat_map)
{
HASH_SEQ_STATUS hash_seq;
GlobalRejectMapEntry *rejectmap_entry;
DiskQuotaActiveTableEntry *active_table_entry;
int num_entries, count = 0;
CdbPgResults cdb_pgresults = {NULL, 0};
StringInfoData rows;
StringInfoData active_oids;
StringInfoData sql;
HASH_SEQ_STATUS hash_seq;
GlobalRejectMapEntry *rejectmap_entry;
ActiveTableEntryCombined *active_table_entry;
int num_entries, count = 0;
CdbPgResults cdb_pgresults = {NULL, 0};
StringInfoData rows;
StringInfoData active_oids;
StringInfoData sql;

initStringInfo(&rows);
initStringInfo(&active_oids);
Expand Down

0 comments on commit 24546b2

Please sign in to comment.