Not so fast . . . kindly please do not automatically presume that I know not whereof I speak, and thus to dismiss what <<I>&;gt; say on just that reason.   :-)

In the instant case, a local timer, running for so-many milliseconds locally, is being set against a timeout of a remote TCP/IP exchange.   Necessarily, a timeout against a remote machine will require some additional milliseconds to resolve ... milliseconds spent informing the client that a timeout has, in fact, occurred.   The AnyEvent authors were obviously very sensitive to such issues, and they actually spent quite a bit of time in their documentation discussing it.   It appears to me that the OP’s test-case does not take these issues into proper account, thereby leading to the (erroneous?) observation of “even/odd.”

Basically, “it will take you somewhat-more than 200 milliseconds to conclude that a 200-millisecond timeout has timed-out.”   (And, as Anonymous Monk originally observed, it is actually quite unpredictable just how many “more than” milliseconds, might be required.)   A strictly-local timer knows nothing of real-world network delays, whereas a valid conclusion cannot ignore them.   C’est la guerre ...

EDIT:   Kindly now draw your attention to what krambambuli subsequently posted.   I think he hit an important nail on the head here ... and something I did not know ... to decisively explain the “even/odd” behavior.

I have made new test-code for check pure Timers. And they behave like I expect and everything is clear for me...
So It's seems that problem from my 1 test-code is related to network "magic".

test-code 2: pure timers

#!/usr/bin/perl

use common::sense;
use AnyEvent;
use Data::Dumper;


sub do {
    my $iter = shift;
    my $cv = AnyEvent->condvar; 
    my ($n, $t, $d, $e);
    ($n, $t) = (AnyEvent->now,  AnyEvent->time);
    $d = $t-$n; 
    say "$iter start timer diff = $d, time = $t, now = $n";
    my $timer = AnyEvent->timer(
        after => 2,
        cb    => sub { warn "Timer\n"; $cv->send },
    );

    $cv->recv;

    ($n, $t, $e) = (AnyEvent->now,  AnyEvent->time, AnyEvent->time - $
+t);
    $d = $t-$n; 
    say "$iter got timer diff = $d, elapsed = $e, time = $t, now = $n"
+;
}

for (1..4) {
    &do($_);
    sleep(5);
    warn "\n";
}
[download]

And result:

1 start timer diff = 0.000634908676147461, time = 1427985442.36442, no
+w = 1427985442.36378
Timer
1 got timer diff = 7.31945037841797e-05, elapsed = 2.00119400024414, t
+ime = 1427985444.36561, now = 1427985444.36554

2 start timer diff = 5.00071620941162, time = 1427985449.36625, now = 
+1427985444.36554
Timer
2 got timer diff = 0.000667810440063477, elapsed = 0.00254702568054199
+, time = 1427985449.3688, now = 1427985449.36813

3 start timer diff = 5.00153064727783, time = 1427985454.36966, now = 
+1427985449.36813
Timer
3 got timer diff = 4.67300415039062e-05, elapsed = 0.00129199028015137
+, time = 1427985454.37095, now = 1427985454.3709

4 start timer diff = 5.00066781044006, time = 1427985459.37157, now = 
+1427985454.3709
Timer
4 got timer diff = 0.000627279281616211, elapsed = 0.00244379043579102
+, time = 1427985459.37401, now = 1427985459.37338
[download]

1 - timer invoked exactly after 2 sec of real time
2 - timer set when internal time and real time has 5 sec lag. So timer invoked immediately.
3..4 like 2 iteration

But Im still interested in my first test-code... just for self-education.

Not so fast . . . kindly please do not automatically presume that I know not whereof I speak, and thus to dismiss what say on just that reason. :-)

Automatically? I quoted the baloney line .. timing issues (fuzziness) do not mean race condition they mean timing issues ... doesn't matter at what rate you blink if you've got baloney glasses