Long time to finish convergence

[Wayne Pfeiffer]

Hi,

I am analyzing some very large alignments from which I want to get trees using search1 with a partial tree constraint and 64 threads and 1 TB of memory. After four days, the first analysis seemed nearly converged three days ago but has been taking a very long time to finish. Here is the output in the log file for the last two restarts after two days each:

[24:28:45 -465221416.168404] SLOW spr round 1 (radius: 5)
[29:42:51 -465184805.194529] SLOW spr round 2 (radius: 5)
[34:03:55 -465184247.464055] SLOW spr round 3 (radius: 5)
[38:20:57 -465184200.848865] SLOW spr round 4 (radius: 5)
[42:36:23 -465184200.848060] SLOW spr round 5 (radius: 10)
[00:00:15 -465184200.848060] SLOW spr round 5 (radius: 10)
[06:29:18 -465183591.293159] SLOW spr round 6 (radius: 5)
[12:55:44 -465183591.289291] SLOW spr round 7 (radius: 10)
[21:11:06 -465183591.288691] SLOW spr round 8 (radius: 15)
[37:52:54 -465183591.288567] SLOW spr round 9 (radius: 20)
[00:00:10 -465183591.288567] SLOW spr round 9 (radius: 20)
[21:32:20 -465183591.288542] SLOW spr round 10 (radius: 25)

The searches after SLOW spr round 6 seem unlikely to have made much difference in the tree.

* Is there some way to relax the convergence criterion to get this analysis to finish faster?

Thanks,

Wayne

[Oleksiy Kozlov]

Hi Wayne,

you can try to increase loglh epsilon, e.g. --lh-epsion 1000

Alternatively, you can try our new adaptive search heuristic which should converge (much) faster in
such cases:

https://github.com/amkozlov/raxml-ng/releases/tag/2.0-beta2

Best,
Oleksiy

> --
> You received this message because you are subscribed to the Google Groups "raxml" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to
> raxml+un...@googlegroups.com <mailto:raxml+un...@googlegroups.com>.
> To view this discussion visit https://groups.google.com/d/msgid/raxml/7dd9d9c6-ddd1-4d29-
> a5e3-9b004f100199n%40googlegroups.com <https://groups.google.com/d/msgid/raxml/7dd9d9c6-ddd1-4d29-
> a5e3-9b004f100199n%40googlegroups.com?utm_medium=email&utm_source=footer>.

[Pfeiffer, Wayne]

Hi Oleksiy,

Thanks for the suggestions. I will see how much they help.

Incidentally, the search listed in my previous email message finished after a total of 187 hours, or nearly 8 days, including restarts. A search without the partial tree constraint seems like it might take several days longer.

Best regards,

Wayne

On Aug 18, 2025, at 2:29 AM, Oleksiy Kozlov <alexei...@gmail.com> wrote:

Hi Wayne,

you can try to increase loglh epsilon, e.g.  --lh-epsion 1000

Alternatively, you can try our new adaptive search heuristic which should converge (much) faster in such cases:

https://urldefense.com/v3/__https://github.com/amkozlov/raxml-ng/releases/tag/2.0-beta2__;!!Mih3wA!HFPJPj8HPXZFOOMTurGnP47WzBVRdaFM_4K9vJFwsrz0nxXbojBgM6E8ipLVX3fR9lwa2TZxw3smnkLPYhcI0_br$

Best,
Oleksiy

On 17.08.25 00:46, Wayne Pfeiffer wrote:

Hi,
I am analyzing some very large alignments from which I want to get trees using search1 with a partial tree constraint and 64 threads and 1 TB of memory. After four days, the first analysis seemed nearly converged three days ago but has been taking a very long time to finish. Here is the output in the log file for the last two restarts after two days each:
[24:28:45 -465221416.168404] SLOW spr round 1 (radius: 5)
[29:42:51 -465184805.194529] SLOW spr round 2 (radius: 5)
[34:03:55 -465184247.464055] SLOW spr round 3 (radius: 5)
[38:20:57 -465184200.848865] SLOW spr round 4 (radius: 5)
[42:36:23 -465184200.848060] SLOW spr round 5 (radius: 10)
[00:00:15 -465184200.848060] SLOW spr round 5 (radius: 10)
[06:29:18 -465183591.293159] SLOW spr round 6 (radius: 5)
[12:55:44 -465183591.289291] SLOW spr round 7 (radius: 10)
[21:11:06 -465183591.288691] SLOW spr round 8 (radius: 15)
[37:52:54 -465183591.288567] SLOW spr round 9 (radius: 20)
[00:00:10 -465183591.288567] SLOW spr round 9 (radius: 20)
[21:32:20 -465183591.288542] SLOW spr round 10 (radius: 25)
The searches after SLOW spr round 6 seem unlikely to have made much difference in the tree.
* Is there some way to relax the convergence criterion to get this analysis to finish faster?
Thanks,
Wayne
-- 
You received this message because you are subscribed to the Google Groups "raxml" group.
To unsubscribe from this group and stop receiving emails from it, send an email to raxml+un...@googlegroups.com <mailto:raxml+un...@googlegroups.com>.

To view this discussion visit https://urldefense.com/v3/__https://groups.google.com/d/msgid/raxml/7dd9d9c6-ddd1-4d29-__;!!Mih3wA!HFPJPj8HPXZFOOMTurGnP47WzBVRdaFM_4K9vJFwsrz0nxXbojBgM6E8ipLVX3fR9lwa2TZxw3smnkLPYioP0Ob0$  a5e3-9b004f100199n%40googlegroups.com<https://urldefense.com/v3/__https://groups.google.com/d/msgid/raxml/7dd9d9c6-ddd1-4d29-__;!!Mih3wA!HFPJPj8HPXZFOOMTurGnP47WzBVRdaFM_4K9vJFwsrz0nxXbojBgM6E8ipLVX3fR9lwa2TZxw3smnkLPYioP0Ob0$  a5e3-9b004f100199n%40googlegroups.com?utm_medium=email&utm_source=footer>.

-- 
You received this message because you are subscribed to the Google Groups "raxml" group.

To unsubscribe from this group and stop receiving emails from it, send an email to raxml+un...@googlegroups.com.
To view this discussion visit https://urldefense.com/v3/__https://groups.google.com/d/msgid/raxml/79877cdd-cb55-4bd0-9060-74d9e201c12e*40gmail.com__;JQ!!Mih3wA!HFPJPj8HPXZFOOMTurGnP47WzBVRdaFM_4K9vJFwsrz0nxXbojBgM6E8ipLVX3fR9lwa2TZxw3smnkLPYp3a_KI4$ .

[Pfeiffer, Wayne]

Hi Oleksiy,

Both of your suggestions speeded up the time to converge :)

For the most part this led to poorer ML scores, though there was one exception where the beta2 version of the code was fastest.

Also, there were some anomalies regarding memory usage and prediction.

I will provide more information in a separate email message to you along with a spreadsheet.

Best regards,

Wayne