RE: BINARIES - San or Local Storage
- From: "Gogala, Mladen" <Mladen.Gogala@xxxxxxxx>
- To: "'oracle-l@xxxxxxxxxxxxx'" <oracle-l@xxxxxxxxxxxxx>
- Date: Fri, 27 Aug 2004 09:29:30 -0400
> -----Original Message-----
> From: Paul Drake [mailto:discgolfdba@xxxxxxxxx]
> Sent: Thursday, August 26, 2004 11:07 PM
> To: oracle-l@xxxxxxxxxxxxx
> Subject: RE: BINARIES - San or Local Storage
>
>
>
> --- "Gogala, Mladen" <Mladen.Gogala@xxxxxxxx> wrote:
>
> > Binaries on the local disk aren't faster then
> > binaries on
> > Symmetrix, as long as Symmetrix isn't configured as
> > RAID-5,
> > please pardon my French. If Symmetrix is configured
> > as RAID-1+0,
> > then it equals or exceeds internal disks in speed
> > and clearly
> > surpasses them in reliability. Problem is that you
> > are not supposed
> > to waste those internal disks either, so you should
> > put there something
> > that can be easily re-created. Oracle executables
> > are (hopefully) not
> > changing on the daily basis. If they are, then you
> > have problem with
> > gremlins who mastered software engineering, which is
> > an especially nasty
> > variety. Internal disks are usually connected to the
> > rest of the system
> > by 320 UW SCSI III with 320 MBit sec transfer rate
>
> Mladen,
>
> Its MegaByte. big M, big B.
> I'm sure that you know this - but - I expect better of
> you.
Actually, I'm still not sure. I know that thruput is calculated as
a number of wires multiplied by frequency. UW SCSI adapters usually
have 32 wires and it is entirely possible that on-board chip runs 10MHZ,
in which case one would see a declared speed of 320 MBit/sec.
> FibreChannel Host Bus Adapters are measured in
> Gigabits per second (2 Gbps).
> Ethernet adapters are measured in Megabit (100 Mbps)
> and Gigabit (1 Gbps).
> Good old SCSI (RAID or otherwise) has always been in
> MB, way back in the days of wide SCSI (20 MB/sec),
> UltraWide SCSI (40 MB/sec), Ultra2Wide SCSI (80
> MB/sec), Ultra 160/m | Ultra3Wide SCSI ... and the
> flavor you mentioned above.
Why would only SCSI adapters be an exception? I am not saying
that you're wrong, but I'm not sure. I'll have to take your
word for it.
> 320 MB/sec is the channel rating speed, nowhere near
> close to the actual sustained transfer rate.
Of course. That speed is calculated as the number of wires in the
channel multiplied by the frequency of the chip that sends the
data to the machine bus and receives it from the machine bus. That
chip can go to the bus only a finite number of times per second, and
to avoid straining the machine bus, those chips run nowhere near the
frequency of the main CPU.
>
> But an 8 drive RAID 10 array could give it a run ...
Yes, if the RAID is properly buffered.
>
> This might seem to be trivial, but it wasn't that long
> ago that a landing craft crashed on the surface of
> Mars ...
>
> > while Symmetrix is
> > usually connected by something called FC/AL (Fiber
> Channel/Arbitrated
> > Loop). Fiber Channel in its latest incarnation surpasses
> > 1GBit/sec, which means
> > that FC controller is significantly faster then
> > SCSI. Disks are usually
> > produced by 3rd party vendor in both cases, so it
> > doesn't really matter.
> > You get 10milisec average access time,
>
> no way. you cannot be serious.
> average access times are much less than that for
> decent server-class drives,
Actually, that is not true. Average seek time for modern drives
is still around 10 miliseconds. That is the state of technology.
> unless your sysadmin gave
> you those 3 x 146 GB drives in an ever so convenient
> RAID 5 array.
>
> > which usually
> > turns into 30 milisec
> > per I/O request, which is excruciatingly slow.
>
> 7 year old hardware, perhaps?
Take a look at your V$FILESTAT and tell me what you see? Lowest numbers
are usually around 10 milisec. If the drive has to do a full seek, and if
you discount the effects of cache, which in some cases you must do, then,
with all the overhead from OS and oracle, it can easily go to 30 milisec.
>
> > Now enters the best EMC
> > cash cow: cache memory.
>
> SCSI RAID controllers have that also.
>
> > It helps significantly
> > reduce average disk access
> > time, but it is not free. As a matter of fact, it is
> > very expensive.
> > Now is the time to stop the story, because we came
> > to the point where I
> > should discuss things like buffer cache efficiency
> > and implementation,
> > priority paging (implemented only by Solaris, of all
> > unixes) and things
> > from the Adrian Cockroft's "Ferrari" book, which
> > definitely outside the
> > scope of this list.
>
> I would hope not.
>
> > Mladen Gogala
>
> If the FibreChannel Host Bus Adapter is saturated
> performing as many IOPS as it can, then its likely
> that you'll have better performance on the internal
> SCSI RAID drives that also have a cache on the ROMB or
> add-in PCI-X controller.
It takes much more traffic and many more disc to staurate 2GBit/sec
adapter then it takes to saturate SCSI. Of course, it all depends on
the particular details.
>
> Anjo Kolk spoke about this at Hotsos.
>
> Think Rambus RAM vs. DDR RAM.
> Think current memory differences of Intel vs. AMD.
>
> Often times, latency effects dominate as opposed to
> bulk, nominal throughput.
> For sheer throughput, you can't beat a cargo ship.
> The 3 month latency might be a show stopper.
Sheer throughput is the name of the game. Cache is here to mitigate the
effects of the latency. At the end of the day, you'll move around more
oil with a 350,000 tons super-tanker then with a speedboat. Tanks on the
coast are here so that one doesn't run out of gas while waiting for the
supertanker. Of course, super-tanker cannot outrun a speed boat.
>
> I think that we need to cover this over a burger and a
> beer, sometime when you can take a 90 minute lunch
> again.
No problem. I agree. I'll be in NYC next week (not for the GOP convention).
----------------------------------------------------------------
Please see the official ORACLE-L FAQ: http://www.orafaq.com
----------------------------------------------------------------
To unsubscribe send email to: oracle-l-request@xxxxxxxxxxxxx
put 'unsubscribe' in the subject line.
--
Archives are at //www.freelists.org/archives/oracle-l/
FAQ is at //www.freelists.org/help/fom-serve/cache/1.html
-----------------------------------------------------------------
Other related posts:
