The 3.03 client... paranoia? :)

[BZ FreeLance]

Ok. I've run the 3.03 client on both my
puters for a few days, and am seeing the 50%
(or so) hit that upgrading from the 3.0
has yeilded.

Here's what I think is going on:
In 3.0, they began using that Japanese fft
routine, and got a large performance increase
so they added Gaussian and Triplet calcs to
gain more information without losing in
performance. This is cool. They went to some
detail in explaining the plusses and features
of the new version.

The 3.0 actually crunched packets faster than
before, so they saw an increase in frequency
of UL/DLs at their site. Coupled with the
normal increase in number of users, they
began feeling the bandwidth pinch.

I think this 3.03 update is simply a return
to the old fft routine, to artificially slow
packet production. The new fft calculations
saved about 50% of the time taken to mow down
a packet, so replacing it with that old fft
calc would add back about 50%, which is about
what I've seen.

Is there any way to check this? If we are
"doing a lot more science", sobeit, I'm there
using the new version. But if I'm simply
going to be penalized in packet production
for installing it early, and there is in fact
no more or better info going back, then I'll
wait until the move is forced on everyone.

Anybody see any detail at all on what "more
science" is being done in the new version?
All I've really seen for detail is concerns
about their ISP and bandwidth...

A bit of a side note... for a few days a
week or so ago, everyone in my group suddenly
began processing packets *very* fast... over
an hour faster. I used to see those once in
a while, every 8 or 10 packets. Then they
dried up for a bit, according to my log, and
then, about the time 3.03 came out, a *lot*
of them came... more than 20 in a row.

I think that was so people would watch the
speed of the new binary, and see that it was
producing packets as fast as in 3.0.... by
offering a lot of the faster processing
packets at once about the time they thought
everyone would be installing the new ver.

This got wordy, sorry, but ... is this just
paranoia? Or could it be, in fact, exactly
what seti@home wants from us... a slower
march of packets/user?

FL

[Pace]

Interesting concept, but I do think they were doing more 'science' - anyone know more?

And, btw, you don't have to press return at the end of the text box, it wraps automatically

Paul.

[Guru]

The bugs seem to be squashed, and now the new version of the S@H GUI is available for download now. There are some changes in the 3.03 version, and some that I really didn't expect. There was one statement from the news from Dec 11, that confused me: "The release of version 3.03 should help alleviate the problem." I scratched my head trying to figure out what they meant by that comment, but from reading the changes in the new client it does become clear what was meant. Here is what was included in the changes:

Windows
Fixed a bug whereby the FFT graph would often not be displayed.
All platforms
Added additional science coverage. We now do a thorough search out to a chirp rate of +- 20 Hz/second. The cost of the additional coverage is that clients will take longer to process a workunit.
Why are we doing this? We have always aimed to balance scientific return with the resources available to the project, the main one being of course our very large and active community of participants. The growth of this community is what enabled us to increase science coverage in version 3.0. Here we were also trying to balance scientific return with the speediness of workunit processing and so we did our most thorough searching out to only +- 10 Hz/second. We sacrificed some science to gain speediness in the client.

Another important resource to the project is our pipe to the Internet. UC Berkeley has been quite generous in allowing us to use the campus ISP, given that SETI@home accounts for 30% of all UC Berkeley Internet traffic! This is an expensive resource ($18,000/month). It has gotten to the point where campus has had to limit our bandwidth usage. We are limited to 30Mbits/s outbound (most of our traffic is outbound). This in turn is resulting in connections getting dropped and generally sluggish server response. We need to back off our bandwidth consumption.

Fortunately, we can do this not by an artificial slowdown but by gaining more science. So, yes, you will see the client slow down, but the server response should pick up. More importantly, you will be doing more science.


Fixed a small bug whereby triplet times were recorded incorrectly in the state file. The times sent back in the result file were correct.

Added a mechanism by which the server can tell the client that it is obsolete and the client will remember that. This will prevent obsolete clients from continually contacting the server.

Changed the alias of the server host. Both the old and new names will be in effect for a while. At some point, we will turn off the old name. The reason is that when we set our server to reject requests from older versions of the client, some of our really old clients (1.x) and some (otherwise great!) third party buffering programs don't have the correct logic to recognize what is happening, and they will keep retrying. By continually contacting our server, they will waste precious bandwidth. Once we turn off the old server name, they will no longer be able to reach the server.
Sorry if it hurts your eyes, but I highlighted the part that many of you wont like in yellow. Version 3.03 will probably take longer to process than the version 3.00 client. Why is it slower? The client is performing more work than before. This isn't a major "addition of science" like the addition of triplet and pulse searches they added in version 3.00, but like performing more searches within a work unit. Why did they do that? Well with the proliferation of large angle range WUs in the past couple of days, it uncovered and magnified the limitations they have with their servers at Berkeley. They push the limits to their servers and bandwidth available to them running the S@H project. Normally, they have some room to spare, but it not that much room. Their two main options to give their servers more room are: 1) Increase server capacity/bandwidth with money they don't have, or 2) Lengthen the processing of a work unit so clients will not connect as quickly, and therefore reduce the load. They chose #2.

The other major change in the client is the changing of the server host. Version 3.03 will be a forced upgrade (eventually). Right now all of the clients resolve to the same address. When the "death" date for old clients comes along, they will shut down the address for the old clients, and switch the address for the version 3.03 clients to a new address. This will prevent all previous clients from connecting the servers to upload/download data. This would effectively force the user to switch clients, or not be able to crunch any more!

Officially it only looks like the Windows, Mac, and Solaris 3.03 clients are ready for download, there are a couple of minor 3.03 clients on the ftp servers. I am not sure when other 3.03 text clients will be available. The other question I know you want to know, is when will this forced upgrade actually be forced? Dunno. It probably won't happen until another month or two down the line. But I will let ya know when I hear it will happen!



------------------
Join the MURC SETI team! | SETI @ MURC

All extremists should be taken out and shot.

[BZ FreeLance]

Cool then I'll let my people know to go
ahead with the upgrade...

I do a little Seti helpsite for the guys in
my group at: http://www.teambattlezone.com/news/setipage1.htm

Certainly not as elaborate as the MURC site,
wish I'd known about this site when it was
first set up It's a good resource!

It never ceases to amaze me, the time a lot
of people put into sites like this one...the people who run the site, and the people who
write into these forums...keep it up guys

FL

[Jorden]

The new client is much slower so Seti can sleep again without server breakdowns

They do it to relief their servers from 50 persons per minute who want a new WU, or so they say... Maybe they need more bandwidth but are afraid to ask

Btw, I don't think Guru wrote the whole post up above ... That's the best English I ever read of him before He must've swallowed an English dictionary (Teasing, Guru !)

Jord.

[This message has been edited by Jorden (edited 24 December 2000).]