Gingyu, This question is a more difficult one not for the technical questions, but by its concepts wit all involved uncertainties.
The maximum of Users is very dependent on a lot of factors. The used Hardware, type of SAS usage, optimizations performance/tuning.
Testing of your implementation how it is performing can be seen as something as kind of regression testing. When planning for a kind of service you should set some kind of benchmark dedicated for your type of service.
Hardware:
You started with an installation and are asking now what it could serve. It is an approach workable in a new environment. In an existing environment you are knowing some of the usage and with that an sizing and than start with the sizing and tuning of the hardware (memory cpu io) with the operating system (OS). Clustering and Grid to be evaluated as options.
SAS usage:
There are several types of usage:
- I interactive The classic information consumers running some known "sas application"
- B batch The classic data processing as executed by an operations support staff. t strategy (Agile waterfall)
- D descriptive adhoc analytics It is the BI part for the next question to answer or self-service BI
- M data/etxt mining
note:
I,B will use a Develop/Test/Acceptance/Production (DTAP) release managemen
D,M will not use the classic DTAP as it developed with production data. There will be a need to segregate
I,D are by nature executed in office hours.
B,M can be scheduled less or more to run outside office hours.
Every usage type can be divided to some ordinals of resources usage:
- light
- medium
- heavy
- extra ordinary
It is about memory/cpu/io and expected runs active at the same moment with a response-time and the DBMS load.
With interactive light-usage there will be a small probability of parallel running. By this you can support a lot of users.
Extra ordinary data mining on the other can be that very demanding that running a couple of uses can exhaust the same machine.
It is good to know about the existence of cgroups - Wikipedia, the free encyclopedia. It is kernel change where it is made possible to do load balancing. It is rather new and not yet accepted as common practice in the Linux world.
By this you have a 4*4 matrix of resource usage with some expectations with assumptions to be filled in.
Knowing the load capacity of your machine/configuration you can set some predictions on the expected load with users.
It will be a continuous measuring and refining to optimize this model. Changing a badly build "sas application" can move that one from heavy to light. To do such things is balancing cost/effort. As you see there will be no fixed defined way how to work at this load cost change performance area.
Resource usage:
Knowing your environment it should be easy to design some tests that will easily overload you machine.
Using Eguide (starting with 5.1) or SAS/connect (MP connect) you can start by yourself a lot of task's in parallel.
Running the same or slightly change code you can get some impression on the performance.
Simple tests:
- Running a proc sort on some faked generated data will cause a heavy load on IO and/or memory using cpu.
having the sizing parameterized you can do this for 1G 10G 20G 50Gb 100Gb sizes dataset. And do this for 1-10 jobs in parallel.
For sure the machine should get overloaded at some moment.
- For the DBMS you can imagine some of the same making difference between reading the DBMS and writing the DBMS
- (more)
You probably did not tuned your SAS options. Sumsize (proc means) memsize bufsize aligniofiles. Some of them will have impact on the bigger data processing.
In a multiuser system it is nor very sensible to give all memory to one process. You can run into http://en.wikipedia.org/wiki/Thrashing_(computer_science).
Connecting to your dbms think on: defer=yes bulkload bufsize. The OS (tcpip) is tuned for 4k block others often for the 64K block sizes.
There is some choice to make, I am used to go for data processing for bigger data, accepting causing some overhead for the smaller ones.
As example a faked dataset, easy to modify to your own ideas. This faked dataset in uncompressed setting is about 1Kb/record. Using the compress=binary is a good tip for datasets having more data fields. If would suggest using an array of 256 character variables 8 bytes long with some random content.
data ivp_run.tst1 ;
length tekst $1024;
do i=1 to &sizenum ;
var1=I; var2=ranuni(10000) ; tekst="äsb 1234567890 qwerty" ; output;
end;
run;
---->-- ja karman --<-----
