[Spread-users] Performance Question
Mike Perik
michaelperik at yahoo.com
Wed Dec 15 17:48:33 EST 2004
Another observation I've noticed is that Spread is
much more bursty than TCP. I would assume that's
because of the token passing?
Mike
--- Mike Perik <michaelperik at yahoo.com> wrote:
> I've done another measurement. For ten minutes I
> publish data on 33 groups. The server publishing
> the
> data timestamps when it sends the message and when
> the
> client receives the data it timestamps. I diff the
> two timestamps for each message and I'm now seeing
> an
> average of >.6 sec delivery time. This is still
> very
> high.
>
> xray_st = spread ticker running on machine xray
> xray_tt = tcp ticker running on machine xray
> frln09_st = client using spread
> frln09_tt = client using tcp (connects to xray_tt)
> Times are in seconds and its the average difference
> between the column and row (ie. time_diff = column -
> row).
>
> frln09_st xray_st frln09_tt xray_tt
>
> frln09_st 0.000000 -0.635060 -0.633197 -0.634968
> xray_st 0.635060 0.000000 0.001863 0.000092
> frln09_tt 0.633197 -0.001863 0.000000 -0.001771
> xray_tt 0.634968 -0.000092 0.001771 0.000000
>
>
> Message rates are:
> Max: 968/s
> Avg: 133/s
>
> All messages are under 64 bytes.
>
>
> I was looking through the mailing list archive and
> found a discussion back on 07/27/2004 about
> adjusting
> the Hurry_timeout to 40 msec which resulted in the
> latency to drop dramaticly. I'm going to try this
> next. Are there any problems with doing this?
>
> Some of the documentation on the website states that
> financial institutions are using Spread. What are
> they using it for? With this kind of latency there
> is no way they can be using if for publishing market
> data.
>
>
> I'd like to see Spread work...
>
> Thanks,
> Mike
>
>
>
> --- Mike Perik <michaelperik at yahoo.com> wrote:
>
> > Yes, they are separated by a switch. I've since
> > removed frln06 from the segment to eliminate it as
> a
> > possible problem. I'm still seeing .7 sec
> > difference
> > in times. I was not expecting this at all and am
> a
> > little disappointed. I understand that you gain
> some
> > things with Spread but .7 sec seems high for just
> > multicasting on a local network.
> >
> > The message size is small, <64 bytes.
> >
> > Are there any optimizations I could do for this
> type
> > of scenario? It's basicly a single, high volume
> > broadcaster and many clients.
> >
> > Would having multiple broadcasters splitting up
> the
> > load allow quicker delivery?
> >
> > How would multiple segments help this? How would
> > that
> > look in the config file?
> >
> > I've heard that OpenAIS may handle this situation
> a
> > little better. Any opinion about that?
> >
> > Thanks,
> > Mike
> >
> > --- Ryan Caudy <rcaudy at gmail.com> wrote:
> >
> > > Is frln06 separated from the others by a router
> or
> > > switch? My reason
> > > for guessing this is the different IP address
> > range
> > > it belongs to. If
> > > so, this may be your problem -- Spread sends
> > > mutlicast messages with a
> > > TTL of 1. You may need to use two
> Spread_Segments
> > > to avoid this
> > > problem. Are the other daemons even finding
> > frln06?
> > >
> > > On a non-lossy network, I would expect to see
> > > slightly higher latency
> > > for messages when comparing Spread and TCP/IP.
> > Note
> > > that, on the
> > > ring, RELIABLE messages are treated the same as
> > FIFO
> > > messages.
> > >
> > > Let me know if this helps.
> > >
> > > Cheers,
> > > Ryan
> > >
> > > On Tue, 14 Dec 2004 10:45:59 -0800 (PST), Mike
> > Perik
> > > <michaelperik at yahoo.com> wrote:
> > > > I've been working on evaluating the
> performance
> > of
> > > > Spread for distributing a Market Data Feed.
> I'm
> > > > subscribing to about 33 high volume stocks and
> > > > broadcasting the information over a 9 node
> > spread
> > > > network. One publisher and 9 clients
> receiving
> > > the
> > > > data.
> > > >
> > > > I'm comparing the times the information is
> > > received on
> > > > the spread system to times taken for a 9
> tcp/ip
> > > > client/server setup.
> > > >
> > > > I'm running both systems on the same machines
> at
> > > the
> > > > same time so I can take the log files and
> strip
> > > out
> > > > the common period of time they ran to do the
> > > > comparison of timestamps.
> > > >
> > > > I'm sending the Spread messages as RELIABLE.
> > > >
> > > > I'm seeing on average that the Spread messages
> > are
> > > > about .7 secs behind the tcp/ip clients.
> > > >
> > > > Has anyone else done similiar tests?
> > > >
> > > > Does this sound right?
> > > >
> > > > I'm a littled surprised by results.
> > > >
> > > > I am seeing a consistant difference in the
> times
> > > > between the first tcp/ip client and the last
> > > tcp/ip
> > > > client which you would expect.
> > > >
> > > > Could the machine frln06 be causing a problem?
> > > >
> > > > The client on that machine does not receive
> any
> > > data.
> > > > I've got my network guy looking into why that
> is
> > > > happening. I believe the router is not
> > configured
> > > > properly.
> > > >
> > > > Spread_Segment 225.0.1.1:5003 {
> > > > frln09 10.0.103.183
> > > > frln11 10.0.103.185
> > > > frln22 10.0.103.100
> > > > frln03 10.0.103.173
> > > > frln16 10.0.103.124
> > > > frln18 10.0.103.130
> > > > wango 10.0.103.102
> > > > frln06 10.0.1.175
> > > > gamma 10.0.103.101
> > > > nero 10.0.103.141
> > > > }
> > > >
> > > > Thanks,
> > > > Mike
> > > >
> > > > __________________________________
> > > > Do you Yahoo!?
> > > > Send a seasonal email greeting and help
> others.
>
=== message truncated ===
__________________________________
Do you Yahoo!?
The all-new My Yahoo! - What will yours do?
http://my.yahoo.com
More information about the Spread-users
mailing list