Category: Test

Maximo with Oracle’s InMemory (Part 2) – Huge Performance Gain

Last week, I played around with Oracle’s new toy: the InMemory feature available in Enterprise Edition. Although it made Maximo runs 1.25x faster, but it didn’t meet my expectation which was from 2x to 5x. This has bothered me for the whole week and I kept thinking about it.

If you’ve read my previous blog post, the one thing I pointed out which could lead to no performance improvement is that I ran the test on a tiny demo version. It has only a few hundred assets and less than a thousand work orders. So, any heavy processes or poorly written queries couldn’t make the database 1 second slower. This week, I set out to do a more elaborate test with a setting that looks more similar to a real production environment.
At first, I setup my system as follows:

  • For Maximo: v7.5 with several industry solutions and add-ons. The data includes 4M work order records, 200k Assets & Locations, 200k Issue Transactions, and 1M Labor Transactions. Total size of the DB is more than 50GB.
  • For DB Server
    • HP EliteBook 840: Core i7-6600U 2.7 Ghz, 16GB RAM, SSD 512. 
    • Oracle 12c Enterprise: 3GB allocated to PGA, 11 GB allocated to SGA (which includes 6GB allocated to InMemory data store)
  • For App Server: Virtual Box running on Mac OS host: Core i5 2.3 GHz, 16GB RAM, SSD 512GB
    • VBox configuration: Windows 2012 x64, 12GB RAM
    • Websphere: Cluster with 3 JVMs x 3GB each.

I got one problem though. The two machines connected wirelessly to each other via a home wifi router which gave a latency of between 5ms to 20ms, sometimes it went up to 100ms or even timeout. This resulted in the system ran with unpredictable performance. In practice, network connection between Application Server and DB Server usually have a latency of  < 1 ms. So I abandoned the idea of having the system runs on two separate boxes, and returned to a configuration with everything runs on one laptop as follows:

  • Database: PGA – 3G, SGA – 8G (5G of which is InMemory)
  • Only one MXServer JVM with 3 GB

I intended to test and compare following operations:

  • Run a Saved Query on Work Order Tracking app
  • Generate 7 PM Work orders at once
  • Change status of 27 Work orders at once
  • Run a simple Report created by “Create Report function” which joins two tables: Work Order and Asset
  • Run a more complex report which joins a few tables including Work Order and Actual Labor and Actual Material

For each of the operations above, I ran test three times to get the average execution time (I ignored the first run as it is usually longer since data hasn’t been populated and cached to memory yet). Below is the result:

For the running saved query, I don’t understand why it took longer to load (result was consistent, not random)

For other expensive operations, Maximo ran 2-3 times faster.

For the more complex report, it never finishes. I tried several times and with the longest one I waited for 50 minutes before killing the process (I checked and Oracle was still processing, not hang). Even after turning on InMemory, it took more than 20 minutes, then I killed it as I was too tired to wait.

For the simple report, it is 5 times faster. If we pay close attention here, the BIRT report window took more than a second to open already, which means DB query returned result almost instantly. This made me believe that we could get a 10-15x performance gain for reports that take several minutes to run.

Since I had to abandon my original setting with two laptops, and also, because I was too lazy to build a new report which will take around 5-20 minutes to run on this environment, I couldn’t test a heavier load scenario with several expensive read/write operations like PM Gen, WO Change Status, and  running several slow reports at once. This kind of load is what we normally see in production and usually slow down the system significantly because operations and queries have to wait for each other to release resource lock. But this situation is exactly where InMemory will make a huge difference.

Anyway, even with this result, I am happy, and from now on, I will be all out selling “memory upgrade” to whomever Maximo users I meet.


An afterthought note: I used to consider DB2 a second choice for Maximo due to the difficulty of finding a good DB2 DBA, and thus, usually only suggested it to small clients with low budget. But now considering the significant cost of Enterprise version of Oracle or SQL if users want to implement this feature, the free bundled DB2 license offered with Maximo is an attractive option. For large enterprise client, I guess I will now present DB2 as an option for them to consider too.

Test Oracle InMemory Database with Maximo

For the last few years, SAP has been pushing hard on its HANA
InMemory data platform and everybody talks about it. For me it makes sense
because SAP’s ERP is such a huge system usually used by super large enterprises
and is both a data intensive and mission critical system.
 
 
Maximo on the other hand is usually much less data intensive
and for most clients I work with in Vietnam, they have small systems with
databases of less than 10-20GB. Thus, I believe InMemory database is not a big
deal for Maximo users. As I recently moved to Australia and got a chance to
work with a much bigger client. Their Maximo runs on a cluster of more than two
dozen JVMs yet somehow is still a bit slow considering the number of active
users that they have. I suspect (since I don’t have visibility to their DB server)
the bottle neck is the database in this case. Besides from the standard
suggestions of looking at disk storage/SAN, network, memory allocation etc., I
also mentioned they can consider implementing InMemory. Then I realized I never
seen it implemented with Maximo, it would be a huge embarrassment if they look
at it and find out that it doesn’t work.
 
This week I have some free time, so I decided to play around
with InMemory database for Maximo to (1) confirm if it is possible and (2) see
if it gives any real performance gain for Maximo.
 
Here is my system configuration:
 
  • Host laptop: Processor Core i7-6600U – 2 cores x
    2.6 Ghz, 16GB RAM, SSD 500GB. Windows 7
  • VMWare: I gave it 10G Ram, 4 cores, Windows 7
  • Database: Oracle 12c R2 x64 Enterprise Edition.
    I gave it 1G for PGA, 4G for SGA (of which, 2G is given to INMEMORY store)
  • Maximo 7.6.0.0 – Demo database. I setup only one
    MXServer and gave it 2G heap size
 
 
 
As you can see, this is very different to a real production
environment and thus the result found in this test may not reflect what you will
find if implemented in production. Some key elements that I can think of that
could lead to differences in results of this setup and real-world production
system include:
 
  • Standard Maximo’s demo database is super small,
    less than 500MB. Thus, InMemory may not lead to any improvement at all
  • Maximo App and Database deployed on the same box.
    Thus, there’s no network latency between App and DB server as usually seen in production
    environment where DB server and App server are placed in two different subnets,
    and can have a firewall in between. This makes resource retrieval process much more
    expensive.
  • This has only one session instead of hundreds if
    not thousands of concurrent sessions as can be seen in production.
  • DB servers on production usually use superfast storage
    systems through SAN/RAID configuration, thus InMemory may not improve
    processing time in that case.
 
 
 
As mentioned, Maximo demo system is very small and most
operations will complete almost instantly (less than .5 second). Thus, I
decided to test two operations:
 
  • Generating Work Order from PM
  • Changing Work Order status
 
The data I chose is the Job Plan: “PMBULKTR – Bulk Trailer PM
Servicing”
, which applies to 5 trailers with Assetnum “44416x” (belong to ‘FLEET
site). The reason I chose this job is because it has 70 tasks. It is a common
problem for Maximo users when changing status of a Work Package with large
number of child WO or tasks is that this operation is very resource intensive
and can take a lot of time to complete. I have seen in many cases, such as in
plant turn-around, changing WO status took so much time that the browser’s
session timed out. So I decided to carry out a test following these steps: Select
the 5 PMs, generate WO at once. Do it three times to measure the time it takes
to complete. After I got 15 new WOs, I move to the Work Order Tracking app, and
select all 15 of them, then try to change status to WAPPR, then to APPR, then
WAPPR again to measure the time it takes each time.
 
After I finished timing the above steps, I will turn on
INMEMORY for the whole ‘MAXIMO’ tablespace by issuing a command: ‘ALTER
TABLESPACE maximo DEFAULT INMEMORY;’
Restart the system, then do the same test again
to measure execution time with INMEMORY enabled.
 
Anyway, my expectation for this configuration would be a 2x
to 5x in term of improved processing speed. The reason is when we built an
add-on for inventory cataloguing, there was a process to compare technical specification
of one item with the spec of thousands of other items to find similarity. This
process took a lot of time to execute. The developer then cached the data in
memory and run the operation with cached data only. This resulted in a boost of
more than 100x in performance. So instead of hours, it takes only a few seconds
to complete. Thus, I think an expectation of 2x to 5x performance gain in this case
is reasonable.
 
Below is the result I got from the test:
 
 
For each operation, I got only about 20% reduction in
execution time. I tried a few other quick tests with other processes while
switching on/off the INMEMORY feature and the result is consistent. Obviously
with this result, my expectation of 2x – 5x improvement is proven to be
unrealistic.
 
Thinking of putting this into practice, if your organization
already have Oracle Enterprise Edition, setting up would be really simple.
Throw in some extra bars of memory, select a few key tables that got queried
the most such as the Work Order – related tables to populate them into INMEMORY
area. If it can give you 20% reduction in execution time, I believe it is still
very attractive to consider.
 
For Microsoft SQL Server, what I found on the web, we need
SQL Server 2016 or 2017 Enterprise Edition which works with Maximo. (I tested
installing Maximo 7.6.0.9 on SQL 2017 and it works well). The free Developer or
Evaluation edition can also be used to test at home. However, I found the configuration
steps seem to be more complex for Maximo as you have to review data structure
to modify/exclude columns with specific datatypes such as XML, BLOB etc. Thus,
if my company is using Maximo with SQL Server, I would think twice about implementing
INMEMORY feature.
 
(Note: I don’t provide detailed steps to configure INMEMORY
feature in Oracle 12c here because it is super simple, all you have to do is
set INMEMORY_SIZE parameter to something other than 0 such as 2G or 4G. Make
sure your SGA is larger than this, e.g. if INMEMORY_SIZE is 2G, SGA should be 3G
or more. Otherwise, you will not be able to start the database service. Once
INMEMORY_SIZE is allocated, issue the ALTER TABLESPACE or ALTER TABLE command
to enable/populate the tablespace or table into INMEMORY area. Also make sure that
you have Oracle 12c Enterprise Edition)