ernestsiu_at_yahoo.com (Ernest Siu) wrote in message news:<2833144d.0402181420.19c30e89_at_posting.google.com>...
> This is getting interesting...
>
> - what is your testing/experimenting experience like? going beyond
> 15km results in magnitude lower performance?
> - how many nodes were you considering?
> - do you find the 'trouble' being private network meta-data exchange?
> - when you say 'too much risk', are you talking about performance? in
> particular, read or write?
> - for your research, what is your proposed storage architecture like?
> SAN? NAS?
> - round trip latency is around 1ms per 100km - what is good latency to
> you?
> - any specific tools/traffic-generator you use?
We did mostly an information search, not a real testing. Our approach
usually is to use some vendor's lab for testing, because we do not
have enough equipment/people/time for this.
Another point is that I didn't say that RAC on greater distances is
impossible, it could probably work up to hundred kilometers. However,
problem is if this is feasible. RAC is usually used for business
critical applications and personally I would never run a business
critical application using technology pushed on its limits. We have
been discouraged by Oracle technicians from ETC from running RAC on
greater distance than 10km. Also a cluster specialist from SUN told us
that on the greater distance we can expect problems and they really
test it. There is not a question if it runs, but there is a question
how it can handle peak loads, how stable it will be and how many
strange bugs you can run into.
So the answers:
1) According to the Oracle experts going beyond 10km is possible, but
not recommended. I really tried to get info from more sources, but it
seems to me like this is the general oficial statement which
represents an Oracle opinion on this.
2) 2
3) I do not understand what you meant. We have leased private
infrastructure based on optical fibre (several gigabites).
4) I've already explained this. In the world of business critical
applications there is no space for bravery. Problems during our RAC
testing were mostly related to performance during heavy write access
using both nodes above the same data, but that's an expected behaviour
if you're dealing with clusters.
5) We're using only SAN. I wouldn't use NAS even in a local
environment.
6) Latency is not only dependent on distance, but also on
infrastructure between. Generally, we have lantencies several ms on
30km. Average latency could be about 1ms on this distance, but a
maximum is higher. And for RAC important is a maximum not an average.
With greater distances you will see not only increase of average, but
also a bigger variance in numbers.
7) For stress testing we're using Compuware's QA Load together with a
script simulating heavy DML load.
--
Dusan Bolek
Email: spambin_at_seznam.cz
Pls add "Not Guilty" to the subject, so your mail will not be burnt.
Received on Thu Feb 19 2004 - 10:33:40 CST
