[
https://issues.apache.org/jira/browse/TEZ-1698?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Gopal V updated TEZ-1698:
-------------------------
Description:
ResourceCalculatorProcessTree scraps all of /proc/ for PIDs which are part of
the current task's process group.
This is mostly wasted in Tez, since unlike YARN which has to do this since it
has the PID for the container-executor process (bash) and has to trace the bash
-> java spawn inheritance.
!ProcfsBasedProcessTree.png!
The latency effect of this is less clearly visible with the profiler turned on
as this is primarily related to rate of syscalls + overhead in the kernel (via
the following codepath in YARN).
!ProcfsFiles.png!
{code}
private List<String> getProcessList() {
String[] processDirs = (new File(procfsDir)).list();
...
for (String dir : processDirs) {
try {
if ((new File(procfsDir, dir)).isDirectory()) {
processList.add(dir);
}
...
public void updateProcessTree() {
if (!pid.equals(deadPid)) {
// Get the list of processes
List<String> processList = getProcessList();
...
for (String proc : processList) {
// Get information for each process
ProcessInfo pInfo = new ProcessInfo(proc);
if (constructProcessInfo(pInfo, procfsDir) != null) {
allProcessInfo.put(proc, pInfo);
if (proc.equals(this.pid)) {
me = pInfo; // cache 'me'
processTree.put(proc, pInfo);
}
}
}
{code}
was:
ResourceCalculatorProcessTree scraps all of /proc/ for PIDs which are part of
the current task's process group.
This is mostly wasted in Tez, since unlike YARN which has to do this since it
has the PID for the container-executor process (bash) and has to trace the bash
-> java spawn inheritance.
!ProcfsBasedProcessTree.png!
The effect of this is less clearly visible with the profiler turned on as this
is primarily related to Syscall overhead in the kernel (via the following
codepath in YARN).
{code}
private List<String> getProcessList() {
String[] processDirs = (new File(procfsDir)).list();
...
for (String dir : processDirs) {
try {
if ((new File(procfsDir, dir)).isDirectory()) {
processList.add(dir);
}
...
public void updateProcessTree() {
if (!pid.equals(deadPid)) {
// Get the list of processes
List<String> processList = getProcessList();
...
for (String proc : processList) {
// Get information for each process
ProcessInfo pInfo = new ProcessInfo(proc);
if (constructProcessInfo(pInfo, procfsDir) != null) {
allProcessInfo.put(proc, pInfo);
if (proc.equals(this.pid)) {
me = pInfo; // cache 'me'
processTree.put(proc, pInfo);
}
}
}
{code}
> Use ResourceCalculatorPlugin instead of ResourceCalculatorProcessTree in Tez
> ----------------------------------------------------------------------------
>
> Key: TEZ-1698
> URL: https://issues.apache.org/jira/browse/TEZ-1698
> Project: Apache Tez
> Issue Type: Bug
> Affects Versions: 0.5.2
> Reporter: Gopal V
> Attachments: ProcfsBasedProcessTree.png, ProcfsFiles.png
>
>
> ResourceCalculatorProcessTree scraps all of /proc/ for PIDs which are part of
> the current task's process group.
> This is mostly wasted in Tez, since unlike YARN which has to do this since it
> has the PID for the container-executor process (bash) and has to trace the
> bash -> java spawn inheritance.
> !ProcfsBasedProcessTree.png!
> The latency effect of this is less clearly visible with the profiler turned
> on as this is primarily related to rate of syscalls + overhead in the kernel
> (via the following codepath in YARN).
> !ProcfsFiles.png!
> {code}
> private List<String> getProcessList() {
> String[] processDirs = (new File(procfsDir)).list();
> ...
> for (String dir : processDirs) {
> try {
> if ((new File(procfsDir, dir)).isDirectory()) {
> processList.add(dir);
> }
> ...
> public void updateProcessTree() {
> if (!pid.equals(deadPid)) {
> // Get the list of processes
> List<String> processList = getProcessList();
> ...
> for (String proc : processList) {
> // Get information for each process
> ProcessInfo pInfo = new ProcessInfo(proc);
> if (constructProcessInfo(pInfo, procfsDir) != null) {
> allProcessInfo.put(proc, pInfo);
> if (proc.equals(this.pid)) {
> me = pInfo; // cache 'me'
> processTree.put(proc, pInfo);
> }
> }
> }
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
