Environment
This was run on Ubuntu 18.04 with Swift for Tensorflow 0.10:
\`\`\`
$ swift --version
Swift version 5.3-dev (LLVM 55d27a5828, Swift 6a5d84ec08)
Target: x86_64-unknown-linux-gnu
$ which swift
/home/xander/swift-tensorflow-RELEASE-0.10-cuda10.2-cudnn7-ubuntu18.04/usr/bin/swift
\`\`\`
I don't believe it's specific to Swift for Tensorflow.
Additional Detail from JIRA
| | |
|------------------|-----------------|
|Votes | 0 |
|Component/s | libdispatch |
|Labels | Bug |
|Assignee | None |
|Priority | Medium |
md5: 5c2364057133aadaf7ecfa69ab055195
Issue Description:
This is based on the Swift Forums discussion here.
Find attached a screenshot of a perf flamegraph that was captured on 8 hours of a Swift server that uses Swift NIO and serial async DispatchQueues. Notice that everything above the `_dispatch_call_block_and_release` is occurring on the serial DispatchQueues. We think those call stacks are not actually executing on NIO's threads, but have been incorrectly assigned to them by perf. It seems like dispatch is missing some debug symbols, which prevents perf getting an accurate handle on how we get into `_dispatch_call_block_and_release`.
This flame graph was collected using the instructions here.
Lukasa
commented
3 years ago
Attachment: Download
Environment
This was run on Ubuntu 18.04 with Swift for Tensorflow 0.10: \`\`\` $ swift --version Swift version 5.3-dev (LLVM 55d27a5828, Swift 6a5d84ec08) Target: x86_64-unknown-linux-gnu $ which swift /home/xander/swift-tensorflow-RELEASE-0.10-cuda10.2-cudnn7-ubuntu18.04/usr/bin/swift \`\`\` I don't believe it's specific to Swift for Tensorflow.Additional Detail from JIRA
| | | |------------------|-----------------| |Votes | 0 | |Component/s | libdispatch | |Labels | Bug | |Assignee | None | |Priority | Medium | md5: 5c2364057133aadaf7ecfa69ab055195Issue Description:
This is based on the Swift Forums discussion here.
Find attached a screenshot of a perf flamegraph that was captured on 8 hours of a Swift server that uses Swift NIO and serial async DispatchQueues. Notice that everything above the `_dispatch_call_block_and_release` is occurring on the serial DispatchQueues. We think those call stacks are not actually executing on NIO's threads, but have been incorrectly assigned to them by perf. It seems like dispatch is missing some debug symbols, which prevents perf getting an accurate handle on how we get into `_dispatch_call_block_and_release`.
This flame graph was collected using the instructions here.
