[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [hobbit] Help! bbtest-net gets http test timing wrong



So I've been trying for a onth to get this working, to no avail, and have pretty much exhausted everything to figure our why Hobbit randomly gets 3-second return times on HTTP tests.

I'm even willing to call it a kernel problem, a problem with select() -- but it happens in multiple reasonably-contemporary kernels.

I've tried:
* CentOS 4.6 my standard build, CentOS 4.6 out-of-box, CentOS 5.1 out of box -- all have same problem. * Disabled ARES. Checked my DNS servers, they are answering fast. Besides, the tests are against IP addresses, not host names.
* Removed all sysctl settings, let them default.  No change.
* Experimented with concurrency settings on [bbnet].  Doesn't help.
* Ran tcpdumps between the web servers and Hobbit server. The tcpdumps indicate the web server is always answering immediately and sending the response... but the FIN packet (when Hobbit completes the test) is delayed 3 seconds.

This issue tends to move around from host to host. It seems to affect Web servers that are sending static HTML pages, and all of them less than 4000 bytes (many about 155 bytes).

As before, testing with other tools on the box show no issues with network connectivity or responses against the same servers.

The /only/ thing that has come close to helping is adding a setsockopt() to bbtest-net, to set the receive bufer to 1024 bytes. This seems to override something that helps the select() call return better somehow. It's not a reliable or even sensible solution. It also does not work on the 2.6.18 kernel on CentOS 5.1...

I'm really stumped and at the end of the rope on this. Has anyone had anything that looks like this?
-Alan

Alan Sparks wrote:
Does exactly the same thing on a fresh install of CentOS 5, x86_64. All built by hand.
-Alan

Alan Sparks wrote:
I see where the problem seems to be occurring. But for my life I can't understand why.

Packet traces from the Hobbit server and the Web servers showing the 3-second delays show that Hobbit connects, and gets an imediate answer from the server (milliseconds). But the servers show that Hobbit does not close the connection (a FIN packets sends/acks) for 3 seconds.

Looking at the bb-network debugging logging, I see that the select() call sleeps for 3 seconds before returning in these cases. So the only conclusion I can arrive at is that select() doesn't return with the active file descriptors on schedule for some bizarre reason.

For a desperation test, I forced the receive buffer on the sockets to a small number (1024 bytes):
                       if (sockok) {
                               int size = 1024;
                               res = setsockopt(nextinqueue->fd,
SOL_SOCKET, SO_RCVBUF, &size, sizeof(size));

This sortof works. the select() no longer hangs, and the HTTP tests start returning "normal"-ish results, i.e. numbers that match curl and wget statistics.

But, it messes with numbers for other Web servers, the ones that return a page significantly larger than 1024 bytes.

Like I said, I just can't get it. Hobbit or CentOS? There's nothing odd about this build, a generic CentOS 4.6 x86_64 build, Hobbit 4.2 with allinone patch, build for x86_64.

Any suggestions at all? If this isn't the right place to ask, where would be? I can't get my hands around why the only thing that I can't get to work here is Hobbit...

Thanks for your indulgence.  I really wish I could fix this.
-Alan


Alan Sparks wrote:
After some Googling, I have added "AcceptFilter http none" directives to the Apache 2.2 servers, which hasn't really helped anything...

Perhaps I should ask: Can anyone verify Hobbit works correctly on a 64-bit system? Not should, but does, on a Centos 4 or RHEL 4 x86_64 install?

I see a lot of debugging trace stuff (dbgprint calls) in the contest and httptest code. Can anyone tell me how to enable it to trace what Hobbit is doing?

Am really at a loss. This can't be rocket science to get it to probe HTTP correctly. But a week later, I still cannot get it to match any other monitoring tool's results.
-Alan

Alan Sparks wrote:
tcpdumps show a couple of interesting points.

1) There are definitely no DNS lookups occurring as a consequence of the Hobbit probes. No port 53 traffic out.

2) The packets from the Hobbit server, and the incoming packets to the Apache server, sometimes look like:

15:20:01.160095 IP (tos 0x0, ttl 62, id 31129, offset 0, flags [DF], proto 6, length: 60) hobbit.45116 > target.http: S [tcp sum ok] 265769416:265769416(0) win 17520 <mss 8760,sackOK,timestamp 143665233 0,nop,wscale 2>

15:20:04.159715 IP (tos 0x0, ttl 62, id 31131, offset 0, flags [DF], proto 6, length: 60) hobbit.45116 > target.http: S [tcp sum ok] 265769416:265769416(0) win 17520 <mss 8760,sackOK,timestamp 143668233 0,nop,wscale 2>

15:20:04.160223 IP (tos 0x0, ttl 62, id 31133, offset 0, flags [DF], proto 6, length: 40) hobbit.45116 > target.http: . [tcp sum ok] 265769417:265769417(0) ack 1051782089 win 17520

So that accounts for three seconds... it appears there are 2 SYN packets, but the first isn't getting processed and there's a 3-second delay to the next SYN (which gets ACKed). I don't know why this happens only with the Hobbit connections... and I don't know why the first SYN seems to be getting ignored. Server is not at all busy.

-Alan
Tim McCloskey wrote:
I get that wget/curl always work. Not sure what resolver settings may be implemented differently for hobbit.

Still thinking this may be unrelated to hobbit (even though wget/curl work fine for you). We have many apache boxes spanning multiple networks running httpd versions 1.3, 2.0 and 2.2 that hobbit(4.2 with allinone patch) likes just fine and reports accurate times (Seconds: 0.nn). We also have fairly proper forward and reverse DNS records for the systems involved.

I can't imagine hobbit parsing the wrong response times, but if that is the case I wonder what external libraries are used (not hobbit provided libs, as ours parse fine and are likely the same as yours).

Anyway, good luck with the tcpdump.

Regards,

Tim





Alan Sparks wrote:
UseCanonicalName is off, and HostNameLookup is off, on every server, regardless of version.
-Alan

Tim McCloskey wrote:
What do you have for
UseCanonicalName
in the apache 2.0 boxes?





To unsubscribe from the hobbit list, send an e-mail to
hobbit-unsubscribe (at) hswn.dk






To unsubscribe from the hobbit list, send an e-mail to
hobbit-unsubscribe (at) hswn.dk






To unsubscribe from the hobbit list, send an e-mail to
hobbit-unsubscribe (at) hswn.dk






To unsubscribe from the hobbit list, send an e-mail to
hobbit-unsubscribe (at) hswn.dk






To unsubscribe from the hobbit list, send an e-mail to
hobbit-unsubscribe (at) hswn.dk