Infrastructure at your Service

Oracle Team

Exadata X-5 Bare Metal vs. OVM performance

By Franck Pachot

.
The Exadata X5 can be installed Bare Metal or Virtualized. The latter one, Oracle VM allows to create multiple clusters on one machine, is more complex for installation and for storage capacity planning. But it allows a small flexibility on options licencing. Those are the real challenges behind the choice. However, when we talk about virtualization, most of questions are about the overhead. Last week, we did some tests on same machine with different configuration, thanks to Arrow Oracle Authorized Solution Center.

Comparison is not easy. Bare Metal has all resources. Virtualized has to distribute resources. And this test is very specific: all I/O hitting the ‘extreme’ flash cache because I don’t expect any virtualization overhead to be in milliseconds. So, don’t expect some universal conclusions from those tests. And don’t hesitate to comment about my way to read those numbers.

CPU

Do not expect a benchmark that shows the maximum capacity of the machine here. I’m comparing a bare metal node with 36 cores with a VM with 4 vCPUS. So I’ll compare a one thread workload only: SLOB with one session and SCALE=100M UPDATE_PCT=0 RUN_TIME=120 WORK_UNIT=64

Bare Metal load profile:

Load Profile Per Second Per Transaction Per Exec Per Call
~~~~~~~~~~~~~~~ --------------- --------------- --------- ---------
DB Time(s): 1.0 30.5 0.00 2.91
DB CPU(s): 1.0 29.0 0.00 2.76
Background CPU(s): 0.0 0.2 0.00 0.00
Redo size (bytes): 14,172.4 432,594.0
Logical read (blocks): 810,244.4 24,731,696.3
Block changes: 41.7 1,271.3
Physical read (blocks): 111.6 3,407.8
Physical write (blocks): 0.0 0.3
Read IO requests: 111.3 3,397.3
Write IO requests: 0.0 0.3
Read IO (MB): 0.9 26.6
Write IO (MB): 0.0 0.0
Executes (SQL): 12,285.1 374,988.5

Virtualized load profile:

Load Profile Per Second Per Transaction Per Exec Per Call
~~~~~~~~~~~~~~~ --------------- --------------- --------- ---------
DB Time(s): 1.0 30.6 0.00 4.37
DB CPU(s): 1.0 29.8 0.00 4.26
Background CPU(s): 0.0 0.2 0.00 0.00
Redo size (bytes): 13,316.5 410,404.0
Logical read (blocks): 848,095.1 26,137,653.8
Block changes: 41.1 1,266.3
Physical read (blocks): 109.1 3,361.3
Physical write (blocks): 0.0 0.3
Read IO requests: 103.8 3,198.5
Write IO requests: 0.0 0.3
Read IO (MB): 0.9 26.3
Write IO (MB): 0.0 0.0
Executes (SQL): 13,051.2 402,228.0

We can say that CPU and RAM performance is similar.

I/O

Now about IOPS on the storage cell flash cache.
I’ll compare SLOB with one session and SCALE=100000M UPDATE_PCT=100 RUN_TIME=120 WORK_UNIT=64

Bare Metal load profile:

Load Profile Per Second Per Transaction Per Exec Per Call
~~~~~~~~~~~~~~~ --------------- --------------- --------- ---------
DB Time(s): 1.0 0.0 0.02 4.06
DB CPU(s): 0.1 0.0 0.00 0.49
Background CPU(s): 0.1 0.0 0.00 0.00
Redo size (bytes): 1,652,624.9 51,700.6
Logical read (blocks): 2,582.2 80.8
Block changes: 4,214.5 131.9
Physical read (blocks): 2,060.6 64.5
Physical write (blocks): 1,818.0 56.9
Read IO requests: 2,051.0 64.2
Write IO requests: 1,738.6 54.4
Read IO (MB): 16.1 0.5
Write IO (MB): 14.2 0.4
Executes (SQL): 66.3 2.1
Rollbacks: 0.0 0.0
Transactions: 32.0

Virtualized load profile:

Load Profile Per Second Per Transaction Per Exec Per Call
~~~~~~~~~~~~~~~ --------------- --------------- --------- ---------
DB Time(s): 1.0 0.0 0.01 3.49
DB CPU(s): 0.3 0.0 0.00 1.01
Background CPU(s): 0.2 0.0 0.00 0.00
Redo size (bytes): 2,796,963.3 51,713.3
Logical read (blocks): 4,226.0 78.1
Block changes: 7,107.0 131.4
Physical read (blocks): 3,470.6 64.2
Physical write (blocks): 3,278.7 60.6
Read IO requests: 3,462.0 64.0
Write IO requests: 3,132.0 57.9
Read IO (MB): 27.1 0.5
Write IO (MB): 25.6 0.5
Executes (SQL): 86.9 1.6
Rollbacks: 0.0 0.0
Transactions: 54.1

In two minutes we did more work here. Timed events show statistics about the ‘cell single block reads’ which are nothing else than ‘db file sequential read’ renamed to look more ‘Exadata’. No SmartScan happens here as they go to buffer cache and we cannot do any filtering for blocks that will be shared with other sessions.

Bare Metal:
Total Wait Wait % DB Wait
Event Waits Time (sec) Avg(ms) time Class
------------------------------ ----------- ---------- ---------- ------ --------
cell single block physical rea 249,854 115.7 0.46 94.9 User I/O
DB CPU 14.6 12.0

Virtualized: Total Wait Wait % DB Wait
Event Waits Time (sec) Avg(ms) time Class
------------------------------ ----------- ---------- ---------- ------ --------
cell single block physical rea 425,071 109.3 0.26 89.4 User I/O
DB CPU 35.2 28.8

Lower latency here on average which explains why we did more work. But no conclusion before we know where the latency comes from. Averages hides the details, and it’s the same with the ‘IO Profile’ section:

Bare Metal
IO Profile Read+Write/Second Read/Second Write/Second
~~~~~~~~~~ ----------------- --------------- ---------------
Total Requests: 3,826.6 2,055.1 1,771.5
Database Requests: 3,789.5 2,051.0 1,738.6
Optimized Requests: 3,720.7 1,985.1 1,735.6
Redo Requests: 32.5 0.0 32.5
Total (MB): 32.0 16.2 15.9
Database (MB): 30.3 16.1 14.2
Optimized Total (MB): 29.3 15.6 13.7
Redo (MB): 1.7 0.0 1.7
Database (blocks): 3,878.6 2,060.6 1,818.0
Via Buffer Cache (blocks): 3,878.6 2,060.6 1,818.0
Direct (blocks): 0.0 0.0 0.0

Virtualized
IO Profile Read+Write/Second Read/Second Write/Second
~~~~~~~~~~ ----------------- --------------- ---------------
Total Requests: 6,652.2 3,467.0 3,185.2
Database Requests: 6,594.0 3,462.0 3,132.0
Optimized Requests: 6,582.7 3,461.2 3,121.5
Redo Requests: 54.7 0.0 54.7
Total (MB): 55.6 27.2 28.4
Database (MB): 52.7 27.1 25.6
Optimized Total (MB): 51.8 27.1 24.6
Redo (MB): 2.8 0.0 2.8
Database (blocks): 6,749.3 3,470.6 3,278.7
Via Buffer Cache (blocks): 6,749.3 3,470.6 3,278.7
Direct (blocks): 0.0 0.0 0.0

and for IO statistics.
Bare Metal:
Reads: Reqs Data Writes: Reqs Data Waits: Avg
Function Name Data per sec per sec Data per sec per sec Count Tm(ms)
--------------- ------- ------- ------- ------- ------- ------- ------- -------
Buffer Cache Re 1.9G 2050.9 16.093M 0M 0.0 0M 250.2K 0.5
DBWR 0M 0.0 0M 1.7G 1740.5 14.216M 0 N/A
LGWR 0M 0.0 0M 201M 32.5 1.648M 3914 0.3
Others 8M 4.1 .066M 1M 0.5 .008M 560 0.0
TOTAL: 1.9G 2055.0 16.159M 1.9G 1773.4 15.872M 254.6K 0.5

Virtualized:
Reads: Reqs Data Writes: Reqs Data Waits: Avg
Function Name Data per sec per sec Data per sec per sec Count Tm(ms)
--------------- ------- ------- ------- ------- ------- ------- ------- -------
Buffer Cache Re 3.3G 3462.7 27.12M 0M 0.0 0M 425.6K 0.3
DBWR 0M 0.0 0M 3.1G 3133.9 25.639M 0 N/A
LGWR 0M 0.0 0M 341M 54.7 2.775M 6665 0.3
Others 10M 5.0 .081M 1M 0.5 .008M 514 0.3
TOTAL: 3.3G 3467.7 27.202M 3.4G 3189.0 28.422M 432.7K 0.3

I’ve put the physical read statistics side-by-side to compare:


BARE METAL VIRTUALIZED
 
Statistic Total per Trans Total per Trans
-------------------------------- ------------------ ------------- ------------------ -------------
cell flash cache read hits 242,142 62.1 425,365 64.0
cell logical write IO requests 5,032 1.3 8,351 1.3
cell overwrites in flash cache 200,897 51.5 937,973 141.1
cell physical IO interconnect by 8,145,832,448 2,089,210.7 14,331,230,720 2,156,044.9
cell writes to flash cache 638,514 163.8 1,149,990 173.0
physical read IO requests 250,168 64.2 425,473 64.0
physical read bytes 2,059,042,816 528,095.1 3,494,084,608 525,663.4
physical read partial requests 4 0.0 0 0.0
physical read requests optimized 242,136 62.1 425,365 64.0
physical read total IO requests 250,671 64.3 426,089 64.1
physical read total bytes 2,067,243,008 530,198.3 3,504,136,192 527,175.6
physical read total bytes optimi 1,993,089,024 511,179.5 3,497,918,464 526,240.2
physical read total multi block 0 0.0 0 0.0
physical reads 251,348 64.5 426,524 64.2
physical reads cache 251,348 64.5 426,524 64.2
physical reads cache prefetch 1,180 0.3 1,051 0.2
physical reads direct 0 0.0 0 0.0
physical reads direct (lob) 0 0.0 0 0.0
physical reads prefetch warmup 1,165 0.3 1,016 0.2
physical write IO requests 212,061 54.4 384,909 57.9
physical write bytes 1,816,551,424 465,901.9 3,300,933,632 496,605.0
physical write requests optimize 211,699 54.3 383,624 57.7
physical write total IO requests 216,077 55.4 391,445 58.9
physical write total bytes 2,026,819,072 519,830.5 3,656,793,600 550,142.0
physical write total bytes optim 1,755,620,352 450,274.5 3,171,875,328 477,189.0
physical write total multi block 531 0.1 942 0.1
physical writes 221,747 56.9 402,946 60.6
physical writes direct 0 0.0 0 0.0
physical writes direct (lob) 0 0.0 0 0.0
physical writes from cache 221,747 56.9 402,946 60.6
physical writes non checkpoint 221,694 56.9 402,922 60.6

We already know that there were more work on the OVM run but comparing the ‘per transaction’ statistics show similar things but a bit more ‘flash cache’ ‘optimized’ I/O in the second run.
Of course, even if it’s the same machine, it has been re-imaged, database re-created, different volume and capacity. So maybe I hit more the cell flash on the second run than on the first one and more reads on spinning disks can explain the difference on single block reads latency.

We need to get beyond the averages with the wait event histograms. They don’t show lower than millisecond in the AWR report (I’ve opened an enhancement request for 12.2 about that) but I collected them from the V$EVENT_HISTOGRAM_MICRO

Bare Metal:
EVENT WAIT_TIME_MICRO WAIT_COUNT WAIT_TIME_FORMAT
---------------------------------------- --------------- ---------- ------------------------------
cell single block physical read 1 0 1 microsecond
cell single block physical read 2 0 2 microseconds
cell single block physical read 4 0 4 microseconds
cell single block physical read 8 0 8 microseconds
cell single block physical read 16 0 16 microseconds
cell single block physical read 32 0 32 microseconds
cell single block physical read 64 0 64 microseconds
cell single block physical read 128 533 128 microseconds
cell single block physical read 256 240142 256 microseconds
cell single block physical read 512 7818 512 microseconds
cell single block physical read 1024 949 1 millisecond
cell single block physical read 2048 491 2 milliseconds
cell single block physical read 4096 1885 4 milliseconds
cell single block physical read 8192 3681 8 milliseconds
cell single block physical read 16384 2562 16 milliseconds
cell single block physical read 32768 257 32 milliseconds
cell single block physical read 65536 52 65 milliseconds
cell single block physical read 131072 3 131 milliseconds
cell single block physical read 262144 0 262 milliseconds
cell single block physical read 524288 1 524 milliseconds

Virtualized:
EVENT WAIT_TIME_MICRO WAIT_COUNT WAIT_TIME_FORMAT
---------------------------------------- --------------- ---------- ------------------------------
cell single block physical read 1 0 1 microsecond
cell single block physical read 2 0 2 microseconds
cell single block physical read 4 0 4 microseconds
cell single block physical read 8 0 8 microseconds
cell single block physical read 16 0 16 microseconds
cell single block physical read 32 0 32 microseconds
cell single block physical read 64 0 64 microseconds
cell single block physical read 128 1 128 microseconds
cell single block physical read 256 322113 256 microseconds
cell single block physical read 512 105055 512 microseconds
cell single block physical read 1024 1822 1 millisecond
cell single block physical read 2048 813 2 milliseconds
cell single block physical read 4096 681 4 milliseconds
cell single block physical read 8192 283 8 milliseconds
cell single block physical read 16384 231 16 milliseconds
cell single block physical read 32768 64 32 milliseconds
cell single block physical read 65536 11 65 milliseconds
cell single block physical read 131072 3 131 milliseconds

In the first run we see more reads around 8ms which confirms the previous guess that we had more flash cache hit on the second run.
The waits between 128 and 512 milliseconds are from the cell flash storage and this is where I want to see if virtualization has an overhead.
I’ve put it in color there where it’s easier to visualize that most of the reads are in the 128-256 range. Bare Metal in blue, OVM in orange.

CaptureX5BMVM

In Bare Metal, most of the reads are faster than 256 microseconds. In virtualized there are some significant reads are above. This may be cause by virtualization but anyway that’s not a big difference. I don’t think that virtualization overhead is an important criteria when choosing how to install your Exadata. Storage capacity planning is the major criteria: consolidate all storage in two diskgroups (DATA and RECO) for all databases, or partition them for each cluster. choice is about manageability and agility in provisioning vs. licence optimization.

Leave a Reply

Oracle Team
Oracle Team