ben-manes / caffeine

94%

master: 100%

LAST BUILD ON BRANCH gradient

branch: gradient

CHANGE BRANCH

x

Reset

• gradient
• 05a040c2478341bab8a58a02b3dc1fe14d626d72
• 2.5.6
• Gil
• adaptive
• async-loader
• ben/issue30
• ben/spacing
• buffer
• cache2k
• clockpro
• codeql
• dependabot/github_actions/JetBrains/qodana-action-2023.1.4
• dependabot/github_actions/JetBrains/qodana-action-61b94e7e3a716dcb9e2030cfd79cd46149d56c26
• dependabot/github_actions/codecov/codecov-action-3.1.3
• dependabot/github_actions/github/codeql-action-2.3.4
• dependabot/gradle/com.gradle.common-custom-user-data-gradle-plugin-1.9
• dependabot/gradle/com.gradle.enterprise-3.14
• dependabot/gradle/de.thetaphi-forbiddenapis-3.6
• dependabot/gradle/pmd-7.0.0-rc4
• expiring_map
• expiry
• gil
• gilga
• guava
• jctools
• juherr
• master
• memory
• orm
• penalties
• solr
• testng
• tmp
• travis
• v1.0
• v1.0.1
• v1.1.0
• v1.2.0
• v1.3.0
• v1.3.1
• v1.3.2
• v1.3.3
• v2.0.0
• v2.0.1
• v2.0.2
• v2.0.3
• v2.1.0
• v2.2.0
• v2.2.2
• v2.2.4
• v2.3.2
• v2.3.4
• v2.5.2
• v2.5.3
• v2.5.4
• v2.5.5
• v2.6.1
• v2.7.0
• v2.8.0
• v2.8.1
• v2.8.2
• v2.8.3
• v2.8.4
• v2.8.5
• v2.dev
• v3.0.5
• v3.0.6
• v3.1.0
• v3.1.1
• v3.1.2
• v3.1.4
• v3.1.5
• v3.1.6
• v3.1.7
• v3.1.8
• v3.dev

Build # 2314

Build Type

push

travis-ci

Committed by ben-manes

Commit Message

Gradient descent optimizers for adaptive tuning

In our paper on adaptive cache policies, we showed how to correct W-TinyLFU
from under performing in recency-biased traces. At its default configuration,
1% window size, it is biased towards frequency. Since the optimal setting is
not known beforehand, we sample the hit rate and dynamically tune to better
values using niave hill climbinb.

The ML community has advanced hill climbing, also known as gradient descent,
for tuning CNN weights. They incorporate momentum, adaptive step sizes, and
bias correction. This discovers the optimal setting faster, better handles
noise and local optimas, and converges.

This change includes SGD with momentum, Adam, Nadam, and AMSGrad; all
popular choices that attempt to improve upon their predecessor. Initial
analysis is extremely promising, e.g. in a recency-biased trace the
default setting has a hit rate of 0.6% and these climbers are optimal
(LRU). In frequency biased, the hit rate did not degrade.

Further analysis is required before incorporating the improvement into
Caffeine.

Run Details

5770 of 6148 relevant lines covered (93.85%)

0.94 hits per line