Replies: 2 comments 1 reply
-
|
Here is a result on the same system with use all warehouses = true (all warehouses = true, stored procedures = true) it is a 3.5% difference between this and use all warehouses = false. The data set is only 400 warehouses (originally we started with 1000 but sysbench took much longer to prepare the schema so we reduced it to save time). Therefore this size of schema will not make a great difference whether using the default 1 warehouse per virtual user or the modified version to use more warehouses to increase I/O. Obviously, it tends to make a bigger difference when you create a much larger data set. Also note as shown at the end of the output HammerDB will purge the history list for more consistency with back to back tests but this is only possible with MariaDB at present. Here is the configuration data and the output from the job showing the warehouse assignment. |
Beta Was this translation helpful? Give feedback.
1 reply
-
|
Hi @sm-shaw ! Thank you for your answer and for all your work! According to the parameters from your test, I found that you ran the test with stored procedures. This is also interesting data, but my point was to add results without stored procedures and with all warehouses, to check clean performance between Sysbench and HammerDB. The list of results is as follows::
Also, if you want to increase I/O but don't want to wait until 1000 warehouses are generated, you can adjust the buffer cache size for InnoDB parameters (innodb_buffer_pool_size from database configuration). Just keep in mind that 1 warehouse is about 100 MB of data, which means that:
I didn't find information about your server memory, but if your memory is more than 100 GB, I suppose the buffer_pool_size parameter is more than 50 GB. This means that there would be no difference between all_warehouses=true and all_warehouses=false, because in both cases, all your data would be in memory. However, if you stay with 400 warehouses, change the buffer cache to around 10 GB, and run with all_warehouses=true, it will be almost the same as if you generated 1600 warehouses. But if you run with 10 GB and with all_warehouses=false, you will get the same results like with 50+ GB buffer cache. Using this trick, you can also run the same test with a 10 GB buffer cache with sysbench and achieve much lower performance because I/O will increase significantly. Sorry to bother you, but I believe that this article can help me convince the customer to switch from sysbench to your tool. |
Beta Was this translation helpful? Give feedback.
-
|
Here is the result for (all warehouses = true, stored procedures = false) Note that as explained comprehensively here https://www.hammerdb.com/blog/uncategorized/why-you-should-benchmark-your-database-using-stored-procedures if you want to compare system capability for running a database then you should use stored procedures, otherwise a big part of your test becomes a network test instead. Here is an archived screenshot with HammerDB as the blue line, sysbench as the orange one.
A key difference is that HammerDB tests Oracle, SQL Server, Db2 and PostgreSQL as well as MySQL and MariaDB on both Windows and Linux so we know what scalability should be achievable for the workload on a particular system as we can try another database engine, when a tool only tests one database on one OS it is difficult to know where the bottlenecks are and therefore how valid the results are. |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
I'm talking about this subject: Comparing HammerDB TPROC-C results with sysbench-tpcc
If I understand correctly, performance between HammerDB and sysbench-tpcc is being checked using only the HammerDB parameter "maria_no_stored_procs." However, you forgot about another parameter, "maria_allwarehouse." As far as I remember, with sysbench-tpcc, you can't choose to use all warehouses or not; it always has "all warehouses = true." But with HammerDB, if "all warehouses = false," your performance will always be better because all your threads read info about warehouses from database caches and not from disks.
So, comparing the following is not fully correct:
You also need to add another test with HammerDB (all warehouses = true, stored procedures = false) to see the gap between sysbench-tpcc and HammerDB.
Beta Was this translation helpful? Give feedback.
All reactions