On 2021/05/11 16:44, Fujii Masao wrote:
>
>
> On 2021/04/28 9:10, Masahiro Ikeda wrote:
>>
>>
>> On 2021/04/27 21:56, Fujii Masao wrote:
>>>
>>>
>>> On 2021/04/26 10:11, Masahiro Ikeda wrote:
>>>>
>>>> First patch has only the changes for pg_stat_wal view.
>>>> ("v6-0001-performance-improvements-of-reporting-wal-stats-without-introducing-a-new-variable.patch")
>>>>
>>>>
>>>
>>> + pgWalUsage.wal_records == prevWalUsage.wal_records &&
>>> + walStats.wal_write == 0 && walStats.wal_sync == 0 &&
>>>> WalStats.m_wal_write should be checked here instead of walStats.wal_write?
>>
>> Thanks! Yes, I'll fix it.
>
> Thanks!
Thanks for your comments!
I fixed them.
>>> Is there really the case where the number of sync is larger than zero when
>>> the number of writes is zero? If not, it's enough to check only the number
>>> of writes?
>>
>> I thought that there is the case if "wal_sync_method" is fdatasync, fsync or
>> fsync_writethrough. The example case is following.
>>
>> (1) backend-1 writes the wal data because wal buffer has no space. But, it
>> doesn't sync the wal data.
>> (2) backend-2 reads data pages. In the execution, it need to write and sync
>> the wal because dirty pages is selected as victim pages. backend-2 need to
>> only sync the wal data because the wal data were already written by
>> backend-1,
>> but they weren't synced.
>
> You're right. So let's leave the check of "m_wal_sync == 0".
OK.
>>> + * wal records weren't generated. So, the counters of 'wal_fpi',
>>> + * 'wal_bytes', 'm_wal_buffers_full' are not updated neither.
>>>
>>> It's better to add the assertion check that confirms
>>> m_wal_buffers_full == 0 whenever wal_records is larger than zero?
>>
>> Sorry, I couldn't understand yet. I thought that m_wal_buffers_full can be
>> larger than 0 if wal_records > 0.
>>
>> Do you suggest that the following assertion is needed?
>>
>> - if (memcmp(&WalStats, &all_zeroes, sizeof(PgStat_MsgWal)) == 0)
>> - return false;
>> + if (pgWalUsage.wal_records == prevWalUsage.wal_records &&
>> + WalStats.m_wal_write == 0 && WalStats.m_wal_sync == 0)
>> + {
>> + Assert(pgWalUsage.wal_fpi == 0 && pgWalUsage.wal_bytes &&
>> + WalStats.m_wal_buffers_full == 0 &&
>> WalStats.m_wal_write_time == 0 &&
>> + WalStats.m_wal_sync_time == 0);
>> + return;
>> + }
>
> I was thinking to add the "Assert(WalStats.m_wal_buffers_full)" as a
> safe-guard
> because only m_wal_buffers_full is incremented in different places where
> wal_records, m_wal_write and m_wal_sync are incremented.
Understood. I added the assertion for m_wal_buffers_full only.
Regards,
--
Masahiro Ikeda
NTT DATA CORPORATION
diff --git a/src/backend/executor/instrument.c b/src/backend/executor/instrument.c
index 237e13361b..75ecd00c23 100644
--- a/src/backend/executor/instrument.c
+++ b/src/backend/executor/instrument.c
@@ -17,6 +17,12 @@
#include "executor/instrument.h"
+/*
+ * Buffer and generated WAL usage counters.
+ *
+ * The counters are accumulated values. There are infrastructures
+ * to add the values and calculate the difference within a specific period.
+ */
BufferUsage pgBufferUsage;
static BufferUsage save_pgBufferUsage;
WalUsage pgWalUsage;
diff --git a/src/backend/postmaster/checkpointer.c b/src/backend/postmaster/checkpointer.c
index e7e6a2a459..1761694a5b 100644
--- a/src/backend/postmaster/checkpointer.c
+++ b/src/backend/postmaster/checkpointer.c
@@ -505,7 +505,7 @@ CheckpointerMain(void)
pgstat_send_bgwriter();
/* Send WAL statistics to the stats collector. */
- pgstat_report_wal();
+ pgstat_send_wal(true);
/*
* If any checkpoint flags have been set, redo the loop to handle the
@@ -583,7 +583,7 @@ HandleCheckpointerInterrupts(void)
BgWriterStats.m_requested_checkpoints++;
ShutdownXLOG(0, 0);
pgstat_send_bgwriter();
- pgstat_report_wal();
+ pgstat_send_wal(true);
/* Normal exit from the checkpointer is here */
proc_exit(0); /* done */
diff --git a/src/backend/postmaster/pgstat.c b/src/backend/postmaster/pgstat.c
index e94f5f55c7..1a1fcc55be 100644
--- a/src/backend/postmaster/pgstat.c
+++ b/src/backend/postmaster/pgstat.c
@@ -133,8 +133,8 @@ PgStat_MsgWal WalStats;
/*
* WAL usage counters saved from pgWALUsage at the previous call to
- * pgstat_report_wal(). This is used to calculate how much WAL usage
- * happens between pgstat_report_wal() calls, by substracting
+ * pgstat_send_wal(). This is used to calculate how much WAL usage
+ * happens between pgstat_send_wal() calls, by substracting
* the previous counters from the current ones.
*/
static WalUsage prevWalUsage;
@@ -852,9 +852,19 @@ pgstat_report_stat(bool disconnect)
TabStatusArray *tsa;
int i;
- /* Don't expend a clock check if nothing to do */
+ /*
+ * Don't expend a clock check if nothing to do.
+ *
+ * Note: regarding to WAL statistics counters, it's not enough to check
+ * only whether the wal record is generated or not. If a read transaction
+ * is executed, wal records aren't nomally generated (although HOT makes
+ * wal records). But, just writes and syncs the wal data may happen when
+ * to flush buffers.
+ */
if ((pgStatTabList == NULL || pgStatTabList->tsa_used == 0) &&
pgStatXactCommit == 0 && pgStatXactRollback == 0 &&
+ pgWalUsage.wal_records == prevWalUsage.wal_records &&
+ WalStats.m_wal_write == 0 && WalStats.m_wal_sync == 0 &&
!have_function_stats && !disconnect)
return;
@@ -948,7 +958,7 @@ pgstat_report_stat(bool disconnect)
pgstat_send_funcstats();
/* Send WAL statistics */
- pgstat_report_wal();
+ pgstat_send_wal(true);
/* Finally send SLRU statistics */
pgstat_send_slru();
@@ -2918,7 +2928,7 @@ void
pgstat_initialize(void)
{
/*
- * Initialize prevWalUsage with pgWalUsage so that pgstat_report_wal() can
+ * Initialize prevWalUsage with pgWalUsage so that pgstat_send_wal() can
* calculate how much pgWalUsage counters are increased by substracting
* prevWalUsage from pgWalUsage.
*/
@@ -3030,44 +3040,6 @@ pgstat_send_bgwriter(void)
MemSet(&BgWriterStats, 0, sizeof(BgWriterStats));
}
-/* ----------
- * pgstat_report_wal() -
- *
- * Calculate how much WAL usage counters are increased and send
- * WAL statistics to the collector.
- *
- * Must be called by processes that generate WAL.
- * ----------
- */
-void
-pgstat_report_wal(void)
-{
- WalUsage walusage;
-
- /*
- * Calculate how much WAL usage counters are increased by substracting the
- * previous counters from the current ones. Fill the results in WAL stats
- * message.
- */
- MemSet(&walusage, 0, sizeof(WalUsage));
- WalUsageAccumDiff(&walusage, &pgWalUsage, &prevWalUsage);
-
- WalStats.m_wal_records = walusage.wal_records;
- WalStats.m_wal_fpi = walusage.wal_fpi;
- WalStats.m_wal_bytes = walusage.wal_bytes;
-
- /*
- * Send WAL stats message to the collector.
- */
- if (!pgstat_send_wal(true))
- return;
-
- /*
- * Save the current counters for the subsequent calculation of WAL usage.
- */
- prevWalUsage = pgWalUsage;
-}
-
/* ----------
* pgstat_send_wal() -
*
@@ -3075,24 +3047,38 @@ pgstat_report_wal(void)
*
* If 'force' is not set, WAL stats message is only sent if enough time has
* passed since last one was sent to reach PGSTAT_STAT_INTERVAL.
- *
- * Return true if the message is sent, and false otherwise.
* ----------
*/
-bool
+void
pgstat_send_wal(bool force)
{
- /* We assume this initializes to zeroes */
- static const PgStat_MsgWal all_zeroes;
static TimestampTz sendTime = 0;
/*
- * This function can be called even if nothing at all has happened. In
- * this case, avoid sending a completely empty message to the stats
- * collector.
+ * First, to check the WAL stats counters were updated.
+ *
+ * Even if the 'force' is true, we don't need to send the stats if the
+ * counters were not updated.
+ *
+ * We can know whether the counters were updated or not to check only
+ * three counters. So, for performance, we don't allocate another memory
+ * spaces and check the all stats like pgstat_send_slru().
+ *
+ * The current 'wal_records' is the same as the previous one means that
+ * wal records weren't generated. So, the counters of 'wal_fpi',
+ * 'wal_bytes', 'm_wal_buffers_full' are not updated neither.
+ *
+ * It's not enough to check the number of generated wal records, for
+ * example the walwriter may write/sync the WAL although it doesn't
+ * generate wal records. 'm_wal_write' and 'm_wal_sync' are zero means the
+ * counters of time spent are zero too.
*/
- if (memcmp(&WalStats, &all_zeroes, sizeof(PgStat_MsgWal)) == 0)
- return false;
+ if (pgWalUsage.wal_records == prevWalUsage.wal_records &&
+ WalStats.m_wal_write == 0 && WalStats.m_wal_sync == 0)
+ {
+ Assert(WalStats.m_wal_buffers_full == 0);
+ return;
+ }
if (!force)
{
@@ -3100,25 +3086,50 @@ pgstat_send_wal(bool force)
/*
* Don't send a message unless it's been at least PGSTAT_STAT_INTERVAL
- * msec since we last sent one.
+ * msec since we last sent one to avoid overloading the stats
+ * collector.
*/
if (!TimestampDifferenceExceeds(sendTime, now, PGSTAT_STAT_INTERVAL))
- return false;
+ return;
sendTime = now;
}
+ /*
+ * Set the counters related to generated WAL data if the counters were
+ * updated.
+ */
+ if (pgWalUsage.wal_records != prevWalUsage.wal_records)
+ {
+ WalUsage walusage;
+
+ /*
+ * Calculate how much WAL usage counters were increased by substracting
+ * the previous counters from the current ones. Fill the results in
+ * WAL stats message.
+ */
+ MemSet(&walusage, 0, sizeof(WalUsage));
+ WalUsageAccumDiff(&walusage, &pgWalUsage, &prevWalUsage);
+
+ WalStats.m_wal_records = walusage.wal_records;
+ WalStats.m_wal_fpi = walusage.wal_fpi;
+ WalStats.m_wal_bytes = walusage.wal_bytes;
+ }
+
/*
* Prepare and send the message
*/
pgstat_setheader(&WalStats.m_hdr, PGSTAT_MTYPE_WAL);
pgstat_send(&WalStats, sizeof(WalStats));
+ /*
+ * Save the current counters for the subsequent calculation of WAL usage.
+ */
+ prevWalUsage = pgWalUsage;
+
/*
* Clear out the statistics buffer, so it can be re-used.
*/
MemSet(&WalStats, 0, sizeof(WalStats));
-
- return true;
}
/* ----------
diff --git a/src/include/executor/instrument.h b/src/include/executor/instrument.h
index aa8eceda5f..9a0d794a1c 100644
--- a/src/include/executor/instrument.h
+++ b/src/include/executor/instrument.h
@@ -16,6 +16,12 @@
#include "portability/instr_time.h"
+/*
+ * The accumulated counters for buffer usage.
+ *
+ * The reason the counters are accumulated values is to avoid unexpected
+ * reset because many callers may use them.
+ */
typedef struct BufferUsage
{
long shared_blks_hit; /* # of shared buffer hits */
@@ -32,6 +38,15 @@ typedef struct BufferUsage
instr_time blk_write_time; /* time spent writing */
} BufferUsage;
+/*
+ * The accumulated counters for generated WAL usage.
+ *
+ * The reason the counters are accumulated values is the same as BufferUsage's one.
+ * And the reason to store only generated WAL usage and doesn't store WAL I/O
+ * activity, is that this is assumed for knowing the WAL usage in per query or
+ * transaction. So, common resources for the cluster like WAL I/O activity is
+ * not stored.
+ */
typedef struct WalUsage
{
long wal_records; /* # of WAL records produced */
diff --git a/src/include/pgstat.h b/src/include/pgstat.h
index 72ff4a06d6..7727e83455 100644
--- a/src/include/pgstat.h
+++ b/src/include/pgstat.h
@@ -1091,8 +1091,7 @@ extern void pgstat_twophase_postabort(TransactionId xid, uint16 info,
extern void pgstat_send_archiver(const char *xlog, bool failed);
extern void pgstat_send_bgwriter(void);
-extern void pgstat_report_wal(void);
-extern bool pgstat_send_wal(bool force);
+extern void pgstat_send_wal(bool force);
/* ----------
* Support functions for the SQL-callable functions to
