Parallel Statement Queuing was introduced in 11.2. Instead of downgrading the requested degree of parallelism (DOP) Oracle will now put your session in a queue until the requested number of slaves become available. Your session would wait for event 'resmgr: pq queued' while waiting in the queue. Like other queues, this is a first in first out queue.
A statement is considered for parallel statement queuing when parallel_degree_policy is set to AUTO. This parameter controls other stuffs as well such as In Memory Parallel Execution and it also turns on auto DOP calculations. So, Oracle will decide for you the degree of parallelism based on the segment size and parameter configurations.
Let's begin with a test case. I'm running Oracle 19c on an Exadata X7 Machine. My parallel related parameters are
parallel_max_servers : 1713
parallel_servers_target : 96
parallel_degree_limit : 128
parallel_degree_policy : AUTO
Parallel_max_servers specifies the maximum number of parallel execution process for an instance. But this parameter is not the upper limit for statement queuing.
A SQL would wait in the queue when
sum of currently active parallel slaves + requested number of slaves exceeds parallel_servers_target.
In my system, parallel_servers_target is set to 96. Imagine I have a query that is currently running with 64 PX slaves. And then another query requests for 64 slaves. Since 64+64 would exceed the parallel_servers_target value of 96, my second query would wait in the queue. As soon as the first query finishes, the second one would come out of the queue and begin execution.
This is why it is important to set the parameter parallel_servers_target to an optimal value. Why does Oracle doesn't consider the parallel max servers instead ? It is because Oracle wants you have two virtual pools of parallel slaves. One pool would serve the queries coming out of auto DOP while others coming from Manual DOP. We will see it shortly.
Let's begin with a test case. I've a table T1 which is query high compressed and the table size is 2.5 TB.
From 1st Session I run this query on T1, Oracle auto calculates the DOP to 64 and SQL begins execution. And while this query is running I open three more sessions and run the same query again
SQL> select /* query 1 */ AVG(COL1) + sum(COL2) from T1
-- SQL Monitor Global Stats section
Global Stats
============================================================================================================================================
| Elapsed | Queuing | Cpu | IO | Other | Fetch | Buffer | Read | Read | Uncompressed | Offload | Offload | Cell |
| Time(s) | Time(s) | Time(s) | Waits(s) | Waits(s) | Calls | Gets | Reqs | Bytes | Bytes | Elig Bytes | Returned Bytes | Offload |
============================================================================================================================================
| 1535 | 0.00 | 368 | 1166 | 0.18 | 1 | 122M | 2M | 2TB | 4TB | 2TB | 130GB | 219.01% |
============================================================================================================================================
Parallel Execution Details (DOP=64 , Servers Allocated=64)
As you can see, Oracle allocates 64 PX slaves. There is no queuing time here because this is the first SQL I'm executing and we haven't yet exhausted the parallel servers target limit of 96.
Now comes the second query, Oracle computes the DOP to be 64 and since this would exceed the target of 96, Oracle places this SQL in the queue. You can see that the queue time for this SQL is 21 seconds.
SQL> select /* query 2 */ AVG(COL1) + sum(COL2) from T1
Global Stats
============================================================================================================================================
| Elapsed | Queuing | Cpu | IO | Other | Fetch | Buffer | Read | Read | Uncompressed | Offload | Offload | Cell |
| Time(s) | Time(s) | Time(s) | Waits(s) | Waits(s) | Calls | Gets | Reqs | Bytes | Bytes | Elig Bytes | Returned Bytes | Offload |
============================================================================================================================================
| 1769 | 21 | 372 | 1375 | 0.75 | 1 | 122M | 2M | 2TB | 4TB | 2TB | 130GB | 219.01% |
============================================================================================================================================
Parallel Execution Details (DOP=64 , Servers Allocated=64)
Now comes my third query. This is interesting, I execute this query with parallel 4 hint. Even though we have hardcoded the DOP to 4, Oracle still considers this as an Auto DOP SQL and this is eligible for parallel statement queuing.
Currently I have 64 slaves active ( first query), second query is in the queue waiting for 64. Currently we have 96-64= 32 free PX slaves. With 32 free slaves, I' requesting just 4. Let's execute the query and see what happens
SQL> select /*+ PARALLEL(4) query 3 */ AVG(COL1) + sum(COL2) from T1
Global Stats
=========================================================================================================================
| Elapsed | Queuing | Cpu | IO | Buffer | Read | Read | Uncompressed | Offload | Offload | Cell |
| Time(s) | Time(s) | Time(s) | Waits(s) | Gets | Reqs | Bytes | Bytes | Elig Bytes | Returned Bytes | Offload |
=========================================================================================================================
| 401 | 8.37 | 319 | 74 | 120M | 2M | 2TB | 4TB | 2TB | 128GB | 219.11% |
=========================================================================================================================
Parallel Execution Details (DOP=4 , Servers Allocated=4)
This SQL too was placed in the queue. Queuing time was 8.37 seconds. So, why was my SQL placed in the queue inspite of the fact that I had 32 PX slaves free and I requested just 4. This is because, our query 3 is behind query 2 in the queue. As I said, its a first in first out queue, so query 2 requesting for 64 PX slaves would get the first priority. Sessions requesting for fewer DOP will not move ahead in the queue.
Now let's run another query. This time with NO_STATEMENT_QUEUING hint. And see what happens.
SQL> select /*+ parallel(4) NO_STATEMENT_QUEUING query 4 */ AVG(COL1) + sum(COL2) from T1
Global Stats
===============================================================================================================
| Elapsed | Cpu | IO | Buffer | Read | Read | Uncompressed | Offload | Offload | Cell |
| Time(s) | Time(s) | Waits(s) | Gets | Reqs | Bytes | Bytes | Elig Bytes | Returned Bytes | Offload |
===============================================================================================================
| 75 | 37 | 38 | 23M | 365K | 356GB | 377GB | 356GB | 5GB | 104.51% |
===============================================================================================================
Parallel Execution Details (DOP=4 , Servers Allocated=4)
So, no queuing time here. This SQL was never placed in the queue. Because we have used the no_statement_queuing hint Oracle does not consider parallel server target as the ceiling value. In fact, now it considers the parallel_max_servers which we have set to 1713.
Basically, Oracle has two separate pools. One pool of 96 PX slaves - this pool is reserved for statements coming from AUTO DOP.
And other pool of 1713-96 = 16717 PX slaves - this pool is reserved for statements coming from manual DOP or for statements coming with no_statement_queuing hint.
CONCLUSION
We need to understand about these virtual pools. If all of my statements are coming thru auto DOP then we must set parallel_servers_target exactly equal to that of parallel_max_servers. And if I have a mixed workload where SQLs come via both auto and manual DOP, then it is best advised to set parallel_servers_target to 50% of max_parallel_servers.
No comments:
Post a Comment