Opened 9 years ago

Closed 9 years ago

Last modified 9 years ago

#1721 closed task (implemented)

Analyze increase in Microsoft Windows relay availability?

Reported by: phobos Owned by: karsten
Priority: Medium Milestone: Deliverable-Sep2010
Component: Metrics/CollecTor Version:
Severity: Keywords:
Cc: Actual Points:
Parent ID: #1753 Points:
Reviewer: Sponsor:

Description

In the past two years, have relays claiming to be microsoft windows had increased uptime? I'm trying to determine if bug #98 is effectively resolved, because the empirical evidence shows that the average uptime of relays has increased over the past 2 years.

Child Tickets

Attachments (4)

relay-uptime.png (28.9 KB) - added by karsten 9 years ago.
Relay uptime by platform
relay-uptime-windows.png (33.2 KB) - added by karsten 9 years ago.
relay-uptime-windows-90perc.png (34.5 KB) - added by karsten 9 years ago.
relay-uptime-ecdf.png (32.8 KB) - added by karsten 9 years ago.
ECDF of uptime sessions

Download all attachments as: .zip

Change History (17)

Changed 9 years ago by karsten

Attachment: relay-uptime.png added

Relay uptime by platform

comment:1 Changed 9 years ago by karsten

The attached file relay-uptime.png shows the average relay uptime in hours by platform from 2008 up to today. There is no significant change in relay uptime for Windows-based relays in the past 2.5 years.

comment:2 Changed 9 years ago by nickm

Can we get this broken down by Windows release (2000, XP, XP Server, Vista, etc) ?

comment:3 Changed 9 years ago by karsten

Sure! Can you let me know what Windows releases you are interested in? Below are the numbers of published server descriptors in July 2010 that have "Windows" in their platform string with all the "Service Pack" and "[workstation]" parts taken out. We should try to combine releases in groups (e.g., is "Very recent version of Windows" == "Vista"?). In particular, averages of node uptime for rarely used releases may not be that meaningful.

43897 on Windows XP
33015 on Very recent version of Windows
17577 on Windows "Longhorn"
7138 on Windows Server 2003
731 on Windows 7
717 on Windows 2000
223 on Windows 98 SE
197 on Windows Vista
13 on Bizarre version of Windows where GetVersionEx doesn't work.

comment:4 Changed 9 years ago by nickm

Hm. It should be possible to figure out what "Very recent version of Windows" is by looking at the rest of the string where it says [major=X,minor=Y]. If major.minor is 6.1, that's Windows 7. If it's 6.0, that's Vista, and so on. From the table in src/common/compat.c :

{ 6, 1, "Windows 7" },
{ 6, 0, "Windows Vista" },
{ 5, 2, "Windows Server 2003" },
{ 5, 1, "Windows XP" },
{ 5, 0, "Windows 2000" },
/* { 4, 0, "Windows NT 4.0" }, */
{ 4, 90, "Windows Me" },
{ 4, 10, "Windows 98" },
/* { 4, 0, "Windows 95" } */
{ 3, 51, "Windows NT 3.51" },
{ 0, 0, NULL }

Also, I am pretty sure that 'Longhorn' is just the older name for 'Vista', so all the Vista and Longhorn entries should be merged.

Also, for Windows XP, since we have so many of those, can we split them up by service pack?

Let's maybe 98SE and Windows 2000 into an 'old versions', or forget about them entirely.

Changed 9 years ago by karsten

Attachment: relay-uptime-windows.png added

comment:5 Changed 9 years ago by karsten

I made a new graph (relay-uptime-windows.png) for five Windows releases (percentages are fractions of published descriptors by Windows relays since January 2008):

  • Windows 7 (8.5 %)
  • Windows Vista (20.6 %)
  • Windows Server 2003 (5.4 %)
  • Windows XP SP3 (35.5 %)
  • Windows XP pre-SP3 (29.6 %)

Most Windows XP pre-SP3 are SP2, but there were not enough SP1 or earlier to justify an own group. Earlier Windows versions (98, 2000) make up only 0.4% of published descriptors, which is why I omitted them.

comment:6 Changed 9 years ago by nickm

Hm. I wonder if this result means that Server 2003 is the most reliable OS, if it just means that people running Server 2003 (the only one apparently that we can identify _as_ a server) decide to reboot or shutdown their OS far less often than people running other versions of Windows.

Would looking at the 80th or 90th percentile uptime rather than the mean uptime answer this question, or am I trying to squeeze information out of this that we just don't have?

comment:7 Changed 9 years ago by arma

My anecdotal evidence is that versions of Windows with the word 'Server' in their name don't suffer from bug 98 in the way that other Windowses do. So I find it perfectly reasonable that Windows Server 2003 has a higher mean uptime.

There are some stable Tor relays that run on it and are up for a long time. There are no XP / etc relays that are stable for a long time.

I'm not sure our data here is good enough to "prove" my anecdote though.

comment:8 Changed 9 years ago by nickm

Milestone: Deliverable-Sep2010
Parent ID: #1753
Summary: Increase in Microsoft Windows relay availability?Analyze increase in Microsoft Windows relay availability?

comment:9 Changed 9 years ago by phobos

Between this analysis and my (#98) testing, it seems we haven't solved the Windows as a relay problems. Feel free to close this. Thanks!

comment:10 Changed 9 years ago by nickm

Karsten -- Before we close this, can we have a look at trends in 90th-percentile uptime by Windows version?

Changed 9 years ago by karsten

comment:11 Changed 9 years ago by karsten

See attached file. Is that what you had in mind?

comment:12 Changed 9 years ago by nickm

Resolution: implemented
Status: newclosed

Yeah; very strange, but it does rule out that one hypothesis I had above. :/ Closing this ticket.

Changed 9 years ago by karsten

Attachment: relay-uptime-ecdf.png added

ECDF of uptime sessions

comment:13 Changed 9 years ago by karsten

Added another graph on uptime sessions, so that we have all of them in one place. This may be useful when integrating similar graphs in the metrics portal.

Note: See TracTickets for help on using tickets.