Jormungandr consumes 100% CPU and crashes

[michaeljfazio]

Describe the bug

After a while running Jormungandr, CPU usage slowly creeps to 100% at which point the node is unable to keep in sync with the network. The only way to recover is to reset the node.

The following diagram illustrates this behaviour:

This is a 2 core node, and 50% CPU utilisation represents and entire core being consumed by Jormungandr (I think because the async thread pool being used is singly threaded in the particular path of code highlighted below).

The sawtooth pattern changes depending on max_connections and gossip period. The period between 16:00 and 08:00 shows a steeper climb to total CPU consumption. Later the climb is not so steep because I have changed max_connections and gossip interval from 512 and 5s to 256 and 10s respectively.

Running a 30 minute profile of Jormungandr (post bootstrap) shows that the majority of CPU usage (greater than 50%) occurs in call to jormungandr::network::send_gossip() function. Within this function, most cycles are consumed deep in the poldercast library during vicinity distance calculation (a merge sort on the total peer list is executed with each iteration).

Mandatory Information

1. jcli --full-version output; jcli 0.8.6 (master-64780cf3, release, linux [x86_64]) - [rustc 1.39.0 (4560ea788 2019-11-04)]
2. jormungandr --full-version output: jormungandr 0.8.6 (performance-e56aada5, release, linux [x86_64]) - [rustc 1.39.0 (4560ea788 2019-11-04)]

Note: Above versions correspond to test branches that im working off, but tip of master is affected in the exact same way.

To Reproduce
Steps to reproduce the behavior:

1. Start Jormungandr with public_address specified and max_connections set to 512.
2. Wait a while....

Expected behavior
Jormungandr runs without failure.

Additional context

It's not clear to me if the root cause here as inefficiencies in the vicinity calculation or a build up of the peer list and subscriptions data structures that are being merge sorted often. Probably a combination of both.

Here is the specific call to sort() which within poldercast lib

One theory is that continuous restarting of nodes currently plaguing the ITN is causing the poldercast node data structures to grow more rapidly than anticipated. Perhaps a bug in the node removal policy is also a factor.

Unfortunately it is quite impossible for me to successfully bootstrap the last few epochs which has not only destroyed sandstone pools profits but also limited my ability to effectively debug this issue further!

[NicolasDP]

Thanks @michaeljfazio definitely some improvements that can be done here. Useful data too.

[Straightpool]

This issue is of very high importance IMHO. Pool operators who have the available funds are arms racing to upgrade their nodes to more and more powerful cores to extend the time until the CPU buildup gets too bad and a restart is inevitable. The only other remedy seems to be limiting max connections and increasing gossip interval.

[AndrewWestberg]

In my own setup (BCSH), I did not see any increase in time until node went to 100% by increasing gossip interval. I saw the same ramp regardless.

However, it seems that incoming connections to the node makes a big difference. If I update the firewall to DENY traffic after my node has gotten a decent number of connected peers, this shifts the burden off of jormungandr and onto the firewall to reject connections. By doing this, I've gotten uptimes regularly over 2 hours instead of 10 minutes when I'm fully open to incoming connections.

I'd prefer this issue get fixed, but in the meantime, I've automated the process of opening and closing the firewall to take the load of jormungandr in my node management software here: https://bitbucket.org/muamw10/jormanager

[dkijania]

Our recent internal soak tests also noticed same behavior:

Quick note about test design:
A single node (stake pool without any trusted peers) is being exercised by moderate amount of transactions (sent via rest api) for couple of days.

The same pattern is observed as in reported issue, however due to isolation of the node from entire network the CPU usage is reaching barely 17% of CPU in 12 hours interval.

[rinor]

@michaeljfazio any improvement small/visible after the changes you made to poldercast (thank you again for those contributions)