ohhhh...
maybe i can find one that has a sleep/pause strategy i can leverage...
~~ #Windows
File system driver samples
2020 December 7 docs.microsoft.com/en-us/windows-…
i can't even remember the name of these things - did we call them repeaters? tack on on to the SAN fabric and it was like adding X km of wire to the route. added latency without adding any compute overhead.
that's basically what i want to do with a filesystem mini filter driver. add at least 10 ms (or some other number) to the erstwhile service time of every single disk IO request that passes through the minifilter driver.
but it can't be accomplished by burning CPU for that 10 ms, or by not yielding the CPU to other threads.
has to be an idle wait, to simulate adding wire time to asynchronous IO.
too much CPU overhead for this, and it invalidates the motivation for the tests.
just like that, i'm probably done thinking about scalabilty strategies for #sqlserver, and back to investigating latch timeouts on very busy systems.
here's the thing. the number of mdmp files for a given large-memory system can be pretty small, especially if an Availability Group is involved.
while a dump is in progress, #sqlserver memory must remain in a constant consistant state. the more memory, the longer the dump may take (there are other un-enumerated factors).
but, Availability Group lease timeout can only be set as high as 100 seconds currently.
484 #sqlserver batch mode on rowstore queries. DOP 8. run consecutively by a single session. each query is a fullscan of a different single table, grouping by a CASE statement on a non-indexed INT/BIGINT/NUMERIC data type column. producing one or two result rows of 1 column each.
based on observation of the first experiment with these queries, in which the queries were ordered alphabetically by the name of the only table in the FROM, i reordered the queries based on decreasing pages:rows ratio. that gives the evident decreasing pages/sec trend :-)
a closeup of the first 10 minutes shows the considerable variation even though the downward trend is evident.
wow
~~
Oracle in Talks to Buy Cerner
An agreement, which could potentially be worth $30 billion, would rank as biggest ever for software giant
2021 December 16 wsj.com/articles/oracl…
i don't know i've ever heard a story about Oracle in talks to buy a company with an ending other than Oracle *buying* that company.
oh. i forgot about the TikTok debacle, mentioned by Financial TImes here.
so that's *one* story about Oracle not buying.
~~
Oracle nears deal to buy health IT company Cerner for $30bn
2021 December 16 ft.com/content/9bf806…
i'll nitpick two details, though :-)
the 512k upper bound for disk IO size isn't properly a #sqlserver limit, although it *is* the limit on most systems. it's the upper bound of the disk adapter.
some disk adapters - most fibre channel HBAs and some others, too - allow configuration of the maximum disk transfer size, and can be set higher than 512kb.
indubitably
~~
"To conduct tactical monitoring, we must also consider the complexities of virtualization, consolidation and concurrency alongside the activities of multiple CPU cores."
"... for capacity planning, this assumption is mostly harmless.
However, this assumption is reckless for tactical monitoring..."