7 Replies Latest reply on Feb 4, 2014 9:31 AM by MKguy

    IOPS meassurements

    RomanB1005 Novice

      Hello

       

       

      At first i want to apologize for an silly question.

       

       

       

      I have small testing lab which is  running on small Synology NAS with two IBM 3560 servers.

      I m trying to resolve my real iops lab performance

       

       

       

      Synology NAS include two mirrored 2 TB WD red edition disks.

      I did deploy two ova templates of  vmware I/O analyzer in order to findout number of IOPs.

       

      Measurements you can see bellow .. (MAX IOPS testing template)

       

      TEST1.jpg

       

       

       

      My question is why this measurements shows around 8000-9000 IOPs when some internet HDD tests of those disks shows around 200 IOPs ... ?

       

       

       

      Thanks in Advance

        • 1. Re: IOPS meassurements
          Abhilashhb Virtuoso
          vExpertVMware Employees

          I'm not too sure but this is what i think.

           

          8000-9000 is the Sum of IOPS for the time interval. And you have ran it for 140 seconds but the time that the values are populated for are 100 seconds. between 20-120

          So 8297/100 is around 82/sec. That's the actual IOPS i think.

          ------------------------------------------------------------------------------------------------------------------------------------
          If you find this or any other answer useful please mark the answer as correct or helpful.
          Abhilash HB
          | Blog : http://vpirate.in | Twitter : @abhilashhb |

          • 2. Re: IOPS meassurements
            JarryG Expert

            I think your ova-templates with I/O analyzer (and esxtop) does not talk to WD "Red" HD directly. They communicate with NAS, which might use its RAM for disk-cache...

            • 3. Re: IOPS meassurements
              MKguy Virtuoso

              Not all IOPS are created equal.

               

              An IO stream can be random or sequential, has a certain read/write ratio, can have IO sizes between 512 Byte and several MB (though that is hardly real-world relevant, up to 128KB is normal).

              All these factors have a huge influence on the "IOs per second" you can measure. Naturally, the smaller the IO size and the more sequential the workload, the higher the IOPS number will be.

               

              When people or vendors say their system supports XXX IOPS, they usually refer to a workload that is 100% random, mainly writes and has a 8-32KB IO size.

               

              It looks like your benchmark is one of these "maximum throughput" tests with a 100% sequential read pattern and a 512 Byte IO size. Doing the math this adds up nicely:

              Your esxtop example shows 8275 read IOs (I suppose the 22 write IOs are generated independently from the benchmark) with a throughput of 4.07MB. This equals 515,7 Byte per IO, pretty close to the minimum IO size of 512 Byte.

               

              The VMware IO analyzer appliance should have patterns with random IO like for Exchange and SQL databases. If you use these IO patterns you will see numbers that are a lot more realistic and lower.

              1 person found this helpful
              • 4. Re: IOPS meassurements
                RomanB1005 Novice

                Thank you for your post.

                So lets take a look on this... ..

                 

                So much bigger iops as declare manufacturer can be caused by ram cache. Right ?

                Regarding max iops pattern , i did try another pre-defined tests and results you can see bellow ...

                 

                Exchange2007.jpgSQL_64k.jpg

                SQL 64K                                                                                                                   Exchange 2007

                 

                Last question ..

                 

                Regarding  last tests , can I run some domain controllers (file server) , mail server and maybe some snmp and vpn server for around 50 users ?

                Is that iops performance enough ?

                 

                I am talking aobout Storage performance ... not network and so on ..

                 

                Silly question , i know ..

                • 5. Re: IOPS meassurements
                  MKguy Virtuoso

                  Your IOPS numbers still look far too unrealistically high for these workloads, considering you only have two spindles in RAID1. I suggest you use IOmeter on a Windows VM with this config file: http://vmktree.org/iometer/OpenPerformanceTest.icf

                   

                  Paste the results on this page: http://vmktree.org/iometer/

                  and compare performance with other setups posted here: New  !! Open unofficial storage performance thread

                  1 person found this helpful
                  • 6. Re: IOPS meassurements
                    RomanB1005 Novice

                    Hi

                     

                    Here are test results ...

                     

                    RESULTS-ADVANCE TEST.jpg

                    I will paste it to , storage performance thread too  ..

                     

                    So what do you think now ?

                     

                    Thanks in Advance ..

                     

                    Roman

                    • 7. Re: IOPS meassurements
                      MKguy Virtuoso

                      Regarding  last tests , can I run some domain controllers (file server) , mail server and maybe some snmp and vpn server for around 50 users ?

                       

                      Is that iops performance enough ?

                      Predicting how such a mixed workload will fare is hard to tell, it's probably mainly depending on what kind of workload these users generate.

                       

                      Looking at the numbers alone, my gut tells me it should be ok for regular users (outside of things like backup hours), but in the end you will need to find out yourself.