Search This Blog

Thursday, August 24, 2017

Effect of MBPS vs Latency

I do *a lot* of performance testing on high performance storage arrays for multiple vendors.  Usually if I'm involved, the client is expecting to put mission critical databases on their new expensive storage, and they need to know it performs well enough to meet their needs. 

So...parsing that out..."meet their needs"...means different things to different people.  Most businesses are cyclical, so the performance they need today is likely not the performance they need at their peek.  For example...Amazon does much more business the day after Thanksgiving than they do in a random day in May.  If you gather the usage stats being used in May and size it appropriately, you're going to get a call in a few months when performance is exposed. 

Before I talk about latency, let me just say AWR does a great job of keeping performance data, if you have your data kept long enough...preferably at least 2 business cycles so you can do comparisons and projections.

This statement will keep AWR data for 3 years, capturing it at an aggressive 15 minute interval:

execute dbms_workload_repository.modify_snapshot_settings (interval => 15,retention => 1576800); that point, see my other post re:gathering IOPS and Throughput requirements.

Anyway, I often have discussions with people who don't understand the effect of latency on OLTP databases.  This is a overly-focused serial example, but its enough to make the point.  Think about this...let's say you have a normal 8K block Oracle database using Netapp or EMC NFS on an active-active 10Gb network.  Let's say your amazing all-SSD storage array is capable of flowing 10Gb between multiple paths.   So...the time to move 8K over a 10Gb pipe is...

(8Kb/s)/(10485760Kb/s)=0.000000762939453125 seconds to copy 8Kb over the 10Gb pipe. the time it passes through your FC network, gets processed by the storage array, gets retrieved from disk, and makes it back to your server can easily exceed a few ms...but for fun let's say we're getting an 8ms response time.  That's .008 seconds.

.008/0.000000762939453125=10,485.76... the effect of latency on your block is 10,485X greater than the effect of throughput.  If your throughput magically got faster but your latency stayed the same...performance wouldn't really improve very much.  If you went from 8ms to 5ms, on the other hand, this would have a huge effect on your database performance.

There's a lot that can affect latency...usually the features in use on the storage array play a big part.  CPU utilization on the storage array can become too high.  This is ultra complicated for the storage array guys to diagnose.  On EMC VMAX3's for example, CPU is allocated to "pools" for different features.  So...even though you may not use eNAS, by default, you allocate a lot of your VMAX CPU cores to it.  When your FC traffic pegs its cores and latency tanks...the administrator may think to look at the CPU utilization and not see an issue...there's free CPU available...just not in the pool used for the FC front end cores, so it creates a bottleneck.  Awesome performance improvements are possible by working closely with your storage vendor to reduce latency during testing...about 6 months ago I worked with a team that achieved improvements by over 50% from the standard VMAX3 as delivered by adjusting those allocations.

All this to say...Latency is very important for common OLTP databases.  Don't ignore throughput, but don't focus on it.

The last secret tweak for improving datapump performance (Part 2)

In my previous post, I mentioned some of the common datapump performance tweaks we see.  In this one, I want to talk about one that's never mentioned, and it might be the best of all.

5.  ADOP - The last tweak...I don't think I've seen any blog posts or Oracle documentation about this as it applies to ADOP-Auto degree of parallelization.  This can be a *huge* 4X or more...improvement on imports, which is typically where most of your datapump time is spent.  This has been around since 11.2, but until recently (12.1,12.2) its been a little difficult to control how parallel things would run at.  To enable it, you simply set:


(which means, if the optimizer thinks this statement will take more than 60 seconds, it will consider parallelizing it)

This is nice because quick queries will run without the overhead of parallelism, and long running queries might find value in parallelism.

Today in 12c, we have the parallel_degree_level parameter, but in 11.2 we could tweak the parallelism by adjusting max_pmbps in sys.resource_io_calibrate$.  From my testing, 200=~parallel 2 or 3.  A SMALLER value increases the amount of parallelism (50=~parallel 20.)  Effectively this gives us the same effect as the new 12c feature...which is to make Oracle make rational decisions on how parallel the auto parallelism should be. 

Datapump is a logical copy (as opposed to a physical backup/restore) so it can't copy the original indexes, it has to rebuild them.  If you have a typical import with 10,000 tables and indexes, the last 100 are big, the last 10 are huge.  Datapump's parallelism will rip through the small objects very quickly, with one process per object.  When the time comes to rebuild the indexes on the huge tables, datapump will again assign one process per index rebuild.  When the create index statement is analyzed by the optimizer, it will create it in parallel based on the algorithm derived from max_pmbps  (even though the create index statement may be parallel 2 or noparallel).  This will save many hours on a large datapump import.  Its crazy to see a serialized create index statement with 50 busy parallel slaves...but that's what can happen.

When its done, the indexes all have the original parallel spec they started with.  Nothing is any different than it would have been if you hadn't used ADOP (other than it was done much, much faster.)

One word of caution:  You have to watch your system resources and parallel limits.  If you have datapump running at parallel 50, that means potentially you'll be rebuilding 50 indexes simultaneously.  If they're each "large" and the optimizer thinks they'll take over [parallel_min_time_threshold] seconds to rebuild, each of them could be built parallel and you could have hundreds (datapump 50 * adop 50) of parallel processes.  This is a wonderful thing if your system can handle it.  Depending on your parallel limit parameters, ADOP may queue the statements until you have enough parallel prevent the system from overloading.  IMHO, that's also a wonderful thing, but it may be unexpected.  The truly unfortunate situation is when you have them too high.  You'll use up your server resources and inefficiently use CPU...and may even swap if you run low on RAM.  So test!

...but that's what dry runs are for.  I hope this last tweak helps you.  I've seen it make miraculous differences meeting otherwise impossible SLA's for datapump export/imports.  Two other posts you may want to read are:
1. Gwen Shapira has the best post on it, IMHO)
2. Kerry Osborne has a nice post on the 12c changes.

Previous post -> The last secret tweak for improving datapump performance (Part 1)
This post -> The last secret tweak for improving datapump performance (Part 2)

The last secret tweak for improving datapump performance (Part 1)

There are 10,000 blog posts and oracle docs on the internets (thank you, Mr Bush) for improving datapump performance.  This is one of the features used very frequently in Oracle shops around the world.  And DBA's are under huge stress to meet impossible downtime SLA's for their export/import. I think they all miss the best tweak (ADOP)...but they basically summarize to one simple fact...outside of parallelism, if you have a well-tuned database on fast storage, there's not a lot more you can do to improve performance more than a few percentage.  The obvious improvements are:

1. Parallelize! To quote Vizzini, "I do not think it means what you think it means."

This will create multiple processes and each will take one object and work with it, each one serialized (usually).  This is great, and if all your objects are equally sized, this is perfect...but that's not typically reality.  Usually you have a few objects that are much bigger than the rest and each of them by default will only get a single process.  This limit really hurts during imports, when you need to rebuild large indexes...more on this later.

Check that its working as expected by hitting control-c and typing in status.  Ideally, you should see all parallel slaves working.  Check not just that they exist, but that they're working (verify you used %U in your dump filename if they aren't.)  ie: DUMPFILE=exaexp%U.dmp PARALLEL=100

2. If you're importing into a new database, you have the flexibility to make some temporary changes to tweak things.  Verify Disk Asynchronous IO is set to true (DISK_ASYNCH_IO=true) and disable all the block verification (DB_BLOCK_CHECK=FALSE, DB_BLOCK_CHECKSUM=FALSE)  These aren't "game changers" but they'll give you 10-20% improvements, depending on how you had them set previously.

3. Memory Settings - Datapump parallelization uses some of the streams API's, and so the streams pool is used.  Make sure you have enough memory for the shared_pool, streams_pool and the db_cache_size parameters.  Consult your gv$streams_pool_advice, gv$shared_pool_advice, gv$db_cache_advice and gv$sga_target_advice  views.  I like to tune it so as the delta in ESTD_DB_TIME_FACTOR from one row to the row below it approaches zero, the corresponding size of the pool is close to 1.  (Any more than that is a waste, any less than that is lost performance.) 

Sometimes you'll see a huge dropoff and its more clear than this example...but you get the idea.  If you're importing into a new database, you'll need to run the import dry run and then check these views to make sure this is tuned well.

158720 0.5 12162648 1.2045
178560 0.5625 11421479 1.1311 0.0734
198400 0.625 10937800 1.0832 0.0479
218240 0.6875 10607608 1.0505 0.0327
238080 0.75 10604578 1.0502 0.0003
257920 0.8125 10375361 1.0275 0.0227
277760 0.875 10222886 1.0124 0.0151
297600 0.9375 10101716 1.0004 0.012
317440 1 10097674 1 0.0004
337280 1.0625 10017905 0.9921 0.0079
357120 1.125 9955300 0.9859 0.0062
376960 1.1875 9908850 0.9813 0.0046
396800 1.25 9905821 0.981 0.0003
416640 1.3125 9870479 0.9775 0.0035
436480 1.375 9844225 0.9749 0.0026
456320 1.4375 9826049 0.9731 0.0018
476160 1.5 9821001 0.9726 0.0005

4. Something often missed when importing into a new database, size your redo logs to be relatively huge.  The redo logs will work like a cache and cycle around.  Eventually if you're adding data extremely fast, the last log will fill and can't switch until the next log is cleared.  "Huge" is relative to the size, speed of your database and hardware.  While you're running your import, select * from v$log and make sure you see at least one "inactive" logs in front of the current log.

The best, virtually unused tweak is in the next post....

This post:  The last secret tweak for improving datapump performance (Part 1)
Next post: The last secret tweak for improving datapump performance (Part 2)