5.0.6: ab-initio reconstruction loop

3D Reconstruction
21

EMail is now also out of the outbox -.-

310GB. I can also share it with “my” service

22

Got it. Thanks.

Perfect. Please let me know the details via email or personal message.

23

Some feedback:

I find that changing the minibatch size helps, probably because the algorithm cannot assign initial particles enough.

Jobs on the older 2080 GPUs are more stable then on the newer 4090 GPUs

24

In our testing, reconstructions based on particles from J13 in the project you shared gave virtually identical, plausible volumes between runs on an rtx4090 and quadro-gp100, respectively.

25

Are you testing with RockyLinux 9 or with another distro?

Thanks

26

The GP100 job ran on ubuntu-26.04 (Note that CryoSPARC v5.0.6 is not currently compatible with the default v7 Linux kernel on ubuntu-26.04).
The RTX4090 job ran on ubuntu-22.04.

27

Any chance to check with RockyLinux 9?

28

I have confirmed the same issue on my PC with at least two datasets.

System Environment:

  • CS Version: 5.0.6
  • OS: Rocky Linux
  • GPU: RTX 4000 Ada x4

Note: The issue is sometimes resolved when stopping other jobs running in parallel.

29

Welcome to the forum @kishi .

Interesting. Please can you post the outputs of these commands

uname -a
cat /sys/kernel/mm/transparent_hugepage/enabled
free -h
30

@wtempel Thank you for reply.

Outputs are below.

uname -a
Linux strbioeng01 5.14.0-687.17.1.el9_8.x86_64 #1 SMP PREEMPT_DYNAMIC Wed Jun 24 13:43:37 UTC 2026 x86_64 x86_64 x86_64 GNU/Linux

cat /sys/kernel/mm/transparent_hugepage/enabled
always madvise [never]

free -h
total used free shared buff/cache available

Mem: 125Gi 21Gi 8.6Gi 3.6Gi 100Gi 104Gi
Swap: 4.0Gi 33Mi 4.0Gi

Best,

31

Thanks @kishi. Did you also observe “blank” volume slices for some classes as @biocit showed in the first post?

32

Yes, I observed blank volume slices as @biocit posted.
The blank volume class was always only one in my cases.

Looking at the logs when it fails (when blank classes appear), the line ‘Current ESS R is 102.88 poses..’ does not seem to have converged sufficiently.
The log below is an example of a successful run, whereas in failed runs, the value stays stuck in the thousands.

Estimating noise model to meet target of 98.17 poses..
Current ESS R is 102.88 poses..
Current sigma is 0.63
New sigma is 0.63

33

You may test this SIGSEGV in core/ioengine/core.so across multiple job types - #4 by hsnyder

I implemented it in my env and asked a user to test multiple ab into jobs at the same time on the same host.

This worked. Either pure luck or the fix helps also for this issue.

34

Thanks, @biocit

As you suggested, I implemented it (CRYOSPARC_IO_URING=false), even when running jobs in parallel, no blank classes have been observed so far.