[email protected] (Timothy Sipples) writes:
> Percentage/ratio cost differences are not *exactly* what drive cost-based
> decisions. To illustrate why, consider these two hypothetical scenarios:
>
> 1. A 1 TB hard disk costs $100 and a 1 TB SSD costs $1000.
> 2. A 1 TB hard disk costs $1 and a 1 TB SSD costs $10.
>
> In both cases the SSD is 10 times more expensive -- 1000%. But in Case #2
> *both* are more affordable -- both experience 99% price drops -- and both
> more easily fit within more budgets. SSD only has to get "affordable
> enough" to take over more use cases from hard disks. Said another way,
> absolute prices matter. Said yet another way, the zero lower bound matters.
>
> A lot of people overlook this reality, but it's a frequent phenomenon in
> many markets.

the PC market with $400-$500 computers would be more price sensitve in
that range ... however there is the issue is how much do the slower or
faster disks affect total system throughput bottleneck.

Said another way, it is the relative system costs that matter (and how
it contributes to total system throughput).

I've periodically mentioned starting to observe in the 70s that disks
were increasingly becoming the bottleneck in overall system throughput.
In the early 80s, I wrote a paper claiming that disk's relative system
throughput had declined by an order of magnitude over the previous 15yrs
... systems got 40-50 times faster while disks got only 4-5 times
faster. Disk division executives took exception and assigned the
division performance group to refute the statement ... after a couple
weeks they came back and effectively said that I had slightly
understated the problem. The analysis is then respun and turns into a
SHARE presentation on how to optimize disk for system throughput.

About the same time there were issues about how datacenter executives
view disk costs based on pure price/megabyte ... they would insist on
filling the (new) 3380 disk drives completely full of data ... or
otherwise they were "wasting" money having half empty 3380s. The issue
was having extra data filling 3380s interferred with disk arm
optimization and accesses for high used data ... degrading overall
system throughput (to save a couple dollars per megabyte on disk they
were willing to sacrifice degraded system throughput of system that
overall ran to tens of millions). There was a semi-facetious proposal
floating at SHARE that IBM announce a special high-performance 3380 that
was much "smaller", "faster" and "more expensive" than standard 3380
disk (with much higher cost/megabyte) that would boost overall system
throughput. In reality it was just a 3880 microcode load that restricted
arm access to only 1/3rd the cylinders. This was something that a
datacenter executive could do all on their own with standard 3380 at
less cost ... but many appeared to be unable to make that leap.

There is recent thread in comp.arch newsgroup about memory being the new
disk and disk being the new tape. If memory & disk access latencies are
measured in number of processor cycles ... the current latency for
memory access measured in number of modern day processor cycles then is
on the same order of 60s disk access latency when measured in 60s
processor cycles.

SSD then might be considered closer to fast tape. But for systems that
are fully utilized and disk throughput is bottleneck factor for overall
system throughput ... then spending thousands of dollars on more
expensive disks might gain several percent increased total system
throughput (for overall datacenter that runs several tens of millions,
especially factoring in total datacenter costs, hardware, cooling,
building, people, etc). Having large controller caches and using memory
for keeping large amount of high use information ... complicates the
analysis.

some recent posts mention SHARE B874 (disk performance group respun analysis)
http://www.garlic.com/~lynn/2011.html#35 CKD DASD
http://www.garlic.com/~lynn/2011.html#61 Speed of Old Hard Disks
http://www.garlic.com/~lynn/2011e.html#1 Multiple Virtual Memory
http://www.garlic.com/~lynn/2011g.html#59 Is the magic and romance killed by 
Windows (and Linux)?
http://www.garlic.com/~lynn/2011p.html#5 Why are organizations sticking with 
mainframes?
http://www.garlic.com/~lynn/2011p.html#32 Has anyone successfully migrated off 
mainframes?
http://www.garlic.com/~lynn/2012b.html#73 Tape vs DASD - Speed/time/CPU 
utilization
http://www.garlic.com/~lynn/2012e.html#39 A bit of IBM System 360 nostalgia
http://www.garlic.com/~lynn/2012o.html#62 ISO documentation of IBM 3375, 3380 
and 3390 track format
http://www.garlic.com/~lynn/2013m.html#72 'Free Unix!': The world-changing 
proclamation made 30 years agotoday
http://www.garlic.com/~lynn/2014b.html#49 Mac at 30: A love/hate relationship 
from the support front
http://www.garlic.com/~lynn/2014l.html#90 What's the difference between doing 
performance in a mainframe environment versus doing in others


-- 
virtualization experience starting Jan1968, online at home since Mar1970

----------------------------------------------------------------------
For IBM-MAIN subscribe / signoff / archive access instructions,
send email to [email protected] with the message: INFO IBM-MAIN

Reply via email to