Switch To CachedIteratorCollection in PeerState

[slumlord]

As per ongoing discussion, I have been testing using CachedIteratorCollection instead of a List in PeerState

I have 2 test runs of 24 hours each - one with a stock 0.9.35 router and another from the i2p.i2p.slumlord branch with the CachedIteratorCollection class and appropriate changes to PeerState to make use of this class.

I focused on 2 main statistics: udp.sendFailed & udp.sendAggressiveFailed

0.9.35:

 udp.sendAggressiveFailed: How many volleys was a packet sent before we gave up
➟ 10 min rate: Average: 11.0; Highest average: 10.999. There were 11 events in this period.
The period ended 2 min ago. (Average event count: 16.821; Events in peak period: 46)
➡ Graph Data - Graph Event Count
➡ Lifetime average value:
 11.0 (2,624 events)

 udp.sendFailed: How many sends a failed message was pushed
➟ 10 min rate: Average: 1.289; Highest average: 1.145. There were 1,225 events in this period.
The period ended 2 min ago. (Average event count: 485.519; Events in peak period: 3485)
➡ Graph Data - Graph Event Count
➡ Lifetime average value: 1.541 (75,741 events)

0.9.35-22-rc-sl

 udp.sendAggressiveFailed: How many volleys was a packet sent before we gave up
➟ 10 min rate: Average: 10.999; Highest average: 11.0. There were 47 events in this period.
The period ended 8 min ago. (Average event count: 21.5; Events in peak period: 50)
➡ Graph Data - Graph Event Count
➡ Lifetime average value:
 11.0 (3,182 events)

 udp.sendFailed: How many sends a failed message was pushed
➟ 10 min rate: Average: 1.512; Highest average: 1.181. There were 1,239 events in this period.
The period ended 8 min ago. (Average event count: 721.554; Events in peak period: 3298)
➡ Graph Data - Graph Event Count
➡ Lifetime average value:
 1.512 (106,790 events)

I have the full /stats page if anyone would like to view other stats.

I have screenshots from the /graphs page as well, these will be attached.

Notes

• Uptime for both version was at least 24 hours before stats & graphs were collected
• Both versions transferred more than 20 GB of data in each direction
• I allowed for automatic floodfill selection, the router was a floodfill when it was running off the i2p.i2p.slumlord branch but not when it was on stock 0.9.35

[slumlord]

udp.sendFailed graph

[slumlord]

udp.sendAggressiveFailed graph

[slumlord]

MemoryUsed? graph

[slumlord]

Bandwidth graph

[slumlord]

ActivePeers? graph

[Zlatin Balevsky]

The stats do not immediately show a problem, but before this gets merged I would like to see a more elaborate description of what the latest problem with add(..) was and a unit test which highlights the issue.

[slumlord]

That was my error, introduced when I was trying to keep the code synced between 3 workspaces and adding changes suggested by zzz:

• Alter the Node object constructor to Node(prev, element) fromNode(prev, element, next)
• Simplify add() so newNode variable declaration/initialization line appears only once

I believe I may have chosen to replicate changes by hand instead of copying the file over from my other workspaces which introduced this error where line 91 read:

final Node<E> newNode = new Node<>(null, element);

instead of

final Node<E> newNode = new Node<>(last, element);

This error is caught by the current set of tests, (which are currently working with JUnit5 - my next task is to tweak the tests to work with the JUnit 4 libraries which we use and check them in so they are run alongside all our other tests) and also results in a lot of NPEs almost immediately.

Also note that I wasn't intending to check this into i2p.i2p just yet but was stumbling around trying to understand how new branches in monotone are set up - I thought I had made a new branch and committed to that but that turned out to be incorrect.

[slumlord]

As per recent discussions, I will run this test one more time with the following changes:

• Ensure that floodfill mode is off
• Collect statistics once more and calculate the following
  • udp.sendFailed per unit of time; per unit of traffic
  • udp.sendAggressiveFailed per unit of time; per unit of traffic

[slumlord]

I have run another live network test of the router with the CachedIteratorCollection class with the following changes to the test:

• Disable floodfill mode
• I decided to run the test for a longer period of time (~72 Hours for stock 0.9.35-23-rc & ~120 Hours for 0.9.35-23-rc-sl 26267b5cf31dfafbdd2c2f8b2e7672afcffaddc9)

Results

|--------------------------+----------------|
| 0.9.35-23-rc             |                |
|--------------------------+----------------|
| udp.sendFailed           | 563,334        |
| udp.sendAggressiveFailed | 13,596         |
|--------------------------+----------------|
| Bandwidth Usage          | 230*2 = 460 GB |
| Uptime                   | ~72 Hours      |
|--------------------------+----------------|

|------------------------------------------+-----------------|
| 0.9.35-23-rc-sl                          |                 |
| 26267b5cf31dfafbdd2c2f8b2e7672afcffaddc9 |                 |
|------------------------------------------+-----------------|
| udp.sendFailed                           | 1,089,969       |
| udp.sendAggressiveFailed                 | 21,993          |
|------------------------------------------+-----------------|
| Bandwidth Usage                          | 600*2 = 1200 GB |
| Uptime                                   | ~120 Hours      |
|------------------------------------------+-----------------|

|------------------------------+-----------------|
| 0.9.35-23-rc                 | 0.9.35-23-rc-sl |
|------------------------------+-----------------+--------------|
| Per Unit of Uptime           |                 | % Difference |
|------------------------------+-----------------+--------------|
| udp.sendFailed               |                 |              |
| 7824.083                     | 9083.075        | +16.09       |
| udp.sendAggressiveFailed     |                 |              |
| 188.83                       | 183.275         | -2.942       |
|------------------------------+-----------------+--------------|
| Per Unit of Data Transferred |                 |              |
|------------------------------+-----------------+--------------|
| udp.sendFailed               |                 |              |
| 1224.639                     | 908.3075        | -25.8305     |
| udp.sendAggressiveFailed     |                 |              |
| 29.556                       | 18.3275         | -37.9905     |
|------------------------------+-----------------+--------------|

I also calculated one more metric: udp.sendFailed/sendAggressiveFailed per unit of uptime per unit of data transferred:

|--------------------------+-----------------+--------------|
| 0.9.35-23-rc             | 0.9.35-23-rc-sl | % Difference |
|--------------------------+-----------------+--------------|
| udp.sendFailed           |                 |              |
| 17.00887                 | 7.56922         | -55.49835    |
|--------------------------+-----------------+--------------|
| udp.sendAggressiveFailed |                 |              |
| 0.4105072                | 0.1527291       | -62.795013   |
|--------------------------+-----------------+--------------|

[slumlord]

Based on discussions at the 26th August 2018 meeting, PeerState in the i2p.i2p branch has been updated to use CachedIteratorCollection from revision fedbe16fb20f7b4f432cdbe04cf6f709257f93cf.
Please test and report any issues found.

Additionally, a set of unit tests for CachedIteratorCollection has been in the i2p.i2p branch since revision 5a69b52dc1480ab57715486420b4a731032ff67f located at router/java/test/junit/net/i2p/router/util/CachedIteratorCollectionTest.java

[slumlord]

Setting this ticket to testing status. Please test and report and issues.

[zzz]

Testing apparently successful