Re: DRBD and Postgres: how to improve the perfomance?

From: Markus Schiltknecht <markus(at)bluegap(dot)ch>
To: "Decibel!" <decibel(at)decibel(dot)org>
Cc: Maila Fatticcioni <mfatticcioni(at)mbigroup(dot)it>, pgsql-performance(at)postgresql(dot)org
Subject: Re: DRBD and Postgres: how to improve the perfomance?
Date: 2007-09-17 15:19:31
Message-ID: 46EE9B03.3000809@bluegap.ch
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-performance

Hi,

Decibel! wrote:
> Actually, in this case, I suspect that latency will be far more critical
> than overall bandwidth. I don't know if it's inherent to Gig-E, but my
> limited experience has been that Gig-E has higher latency than 100mb.

I've been looking for some benchmarks, but it's rather hard to find. It
looks like people are much more concerned about throughput ?!?

However, I'd like to share some of the sites I've found, especially
regarding Fast Ethernet vs. Gigabit Ethernet:

- Ashford Computer Consulting Service benchmarked five different
gigabit ethernet adapters [1], back in 2004. For most cards they
measured between ca. 100 - 150 microseconds for a UDP round trip of a
token, a so called hot potato benchmark. Unfortunately they didn't
compare with Fast Ethernet.

- The NetPIPE project has some of it's measurements at the very bottom
of it's website [2]. Mostly for high speed and low latency links. Again,
Fast Ethernet is missing. The diagram tells the following latencies (in
microseconds):

75 10 Gigabit Ethernet
62 Gigabit Ethernet
8 Myrinet
7.5 Infini Band
4.7 Atoll
4.2 SCI

I've no explanation for the significantly better measure for gigabit
ethernet compared with the above benchmark. From their description I'm
concluding that they also measured a round-trip, but not via UDP.

The bad value for 10 Gigabit Ethernet is due to a poor Intel adapter,
which also has poor throughput. They claim that newer adapters are better.

- Finally, I've found a latency comparison between Fast vs Gigabit
Ethernet, here [3]. Figure 6, in the second third of the page shows a
NetPIPE latency benchmark between Ethernet, Fast Ethernet and Gigabit
Ethernet (additionally ATM and FDDI). It looks like Gigabit Ethernet
features slightly better latency.

From these findings I'm concluding, that commodity Ethernet hardware
has quite similar latencies, no matter if you are using Fast, Gigabit or
10 Gigabit Ethernet. If you really want to have a low latency
interconnect, you need to pay the extra bucks for specialized, low
latency networking hardware (which may still be based on 10GE, see
Myrinet's 10GE adapter).

If you know other resources, I'd be curious to know.

Regards

Markus

[1]: Ashford Computer Consulting Service, GigE benchmarks:
http://www.accs.com/p_and_p/GigaBit/conclusion.html

[2]: NetPIPE website:
http://www.scl.ameslab.gov/netpipe/

[3]: Gigabit Ethernet and Low-Cost Supercomputing
http://www.scl.ameslab.gov/Publications/Gigabit/tr5126.html

In response to

Browse pgsql-performance by date

  From Date Subject
Next Message Carlo Stonebanks 2007-09-17 20:22:50 Query works when kludged, but would prefer "best practice" solution
Previous Message W.Alphonse HAROUNY 2007-09-17 14:40:49 Re: Index usage when bitwise operator is used