My 2 cents. We don't really bother with read_realloc but it's because we're so much more concerned with W (and DEL = reclaim space) type of issues in WAFL. Keeping Free Space (at Aggr level) good is what's important for us.
On 2017-03-08 16:58, Steiner, Jeffrey wrote:
The main improvement you'll see is at the Oracle level. You should get better latency/throughput from direct_path_read and db_file_parallel_read operations. You won't see anything notable in the more common random read latency, but to keep things running cleanly read_realloc is a general best practice when spinning disk is involved.
read_realloc isn't totally free, i.e. it costs *something* to have it running. Extra I/O, which will in some sense put load on disk(s). If that cost is lower than the benefit, is highly dependent on the workload, more precisely the Seq READ portion of it. All workloads have that (more or less), and Seq R always comes from spinning. So for example we have very aggregated random workloads, and lots of HyA with large SSD RGs, as well as some really big FlashCache instances. But even our workload has a portion of Seq R.
If you have, as we do, a HyA "Tier 2" meaning it's based on 7.2K rpm disk, then the Seq R from them, will very likely benefit noticeably from read_realloc. But... it depends if it will be worth it overall.
There's no way to know and very very hard to quantify the result by simply trying it out. It's painstaking work, like a scientific experiment (how do I measure this?) takes weeks to do an experiment.
It doesn't do anything useful with SSD aggregates and I think it's impossible to set in the latest versions of ONTAP anyway.
As a side effect, the overall loading on ONTAP would go down as sequential IO becomes more efficient, but I'm not sure how you'd actually prove that it's occurring unless the currently layout is truly awful.
Ah, well... see above. I'd say it depends if the "overall load" (whatever the definition of that is) of WAFL will go down. read_realloc does have a cost associated with it too I totally agree that to try and prove things around this is dodgy, at least very very hard. And there s definitely no general statement around it, it's each and every man's own system properties and workload dependent
/M
-----Original Message----- From: toasters-bounces@teaparty.net [mailto:toasters-bounces@teaparty.net] On Behalf Of Ehrenwald, Ian Sent: Wednesday, March 08, 2017 4:48 PM To: toasters@teaparty.net Subject: Measurable effects of read_realloc?
Hello Is there a way within an SVM (cDOT 8.3.2P5) to observe the results of setting read_realloc to “on” in Oracle database flexvols? I told my DBAs that I’d be enabling this option on a couple eligible flexvols per day to see if we’d benefit from it, and we all are curious to see if it has benefited us. What sort of measurable things would someone see? Decreased latency and increased throughput after a couple weeks of heavy read load? I have OCUM/OCPM and LogicMonitor to view filer-side stats, and my DBAs have OEM to view stats coming from Oracle. Thanks for any hints.