Page MenuHomeMiraheze

Redesign DNS Geo-location
Open, NormalPublic

Description

With T7750 calling for the decom of cp3 (something I have always advocated since moving across the OVH), we need to re-balance and evaluate how geo-located requests are balanced among datacentres in GDNSD.

This will mostly be based on average latency to a server, I mind expand to a more country based approach if there are significant benefits to splitting a continent in half (mostly around Asia/Oceania?) given we only have two cache PoPs.

There is also a question of whether we would benefit form a third-external PoP, reviewing available options on some large VPS providers, I'll provide an additional analysis of this to see if potential data would back up further investigation into this - or whether investing more in either GB or CA would bring the most benefit.

Key

  • CA = Canada
  • GB = Great Britain
  • SG = Singapore
  • AU = Australia

Event Timeline

John triaged this task as Normal priority.Sun, Sep 5, 14:34
John created this task.
ContinentCountry-> CA-> GB-> SG (hist)->AU (OVH)Best DC?
AFEG14254195298GB
AFZA221144311398GB
AFKE245171231412GB
AFGH175103331344GB
AFTZ216136298403GB
AFMA14263249305GB
AFNG21797396335GB
ASTH25522229122SG
ASIN23114234355SG
ASAE193136266405GB
ASVN28822136132SG
ASID23820225118SG
ASKZ19095252362GB
ASPK274145110199SG
ASMY457294145240SG
ASPH26922966227SG
ASSA18697233258GB
ASKP19626682289SG
ASCN281?66349SG
ASJP16822374217SG
EUGR12755212279GB
EURO11537184271GB
EUIE7816257262GB
EUGB8811171263GB
EUFI11131182300GB
EUTR12551271288GB
EUUA11740177295GB
EUPT11548192282GB
EUSI10638191274GB
EUES10227253252GB
EURU12049207308GB
EUDE9619159251GB
EUNO10236190276GB
EUFR8312151248GB
EUIS12438196294GB
EUIT9919144242GB
NAUS2583236210CA
NACA685235210CA
NAMX71136228197CA
OCAU26026010016AU
OCNZ26030211726AU
SABR133213369345CA
SAAR175221368344CA
SAGT250277490478CA
SAPE107171307277CA
SAUY149227385335CA
SACL138198332306CA

Based on the above, a potential DC layout might be in an ideal world of four PoPs:

  • AF -> GB
  • AS -> SG
  • EU -> GB
  • NA -> CA
  • OC -> AU
  • SA -> CA

In a world of three PoPs (AU):

  • AF -> GB
  • AS -> GB/AU
  • EU -> GB
  • NA -> CA
  • OC -> AU
  • SA -> CA

In a world of three PoPs (SG):

  • AF -> GB
  • AS -> SG
  • EU -> GB
  • NA -> CA
  • OC -> SG
  • SA -> CA

In a world of two PoPs (CA/GB):

  • AF -> GB
  • AS -> CA/GB
  • EU -> GB
  • NA -> CA
  • OC -> GB (CA->GB latency for uncached requests might matter?)
  • SA -> CA

In the above scenarios, I'm looking at potential user/population counts each PoP would be responsible for based on a 24 hour period of rough access.

Data based on Sunday, August 8th 2021. This was our most recent user peak (169k unique).

AF - 1k
AS - 48k
EU - 45k
NA - 64k
OC - 5k
SA - 6k

4PoP World:

  • AU - 5k (3%)
  • CA - 70k (42%)
  • GB - 46k (27%)
  • SG - 48k (28%)

3PoP AU World:

  • AU - 48k (29%)
  • CA - 70k (40%)
  • GB - 51k (30%)

2PoP World:

  • CA - 94k (56%)
  • GB - 75k (44%)

Interestingly enough, in all scenarios - our smallest PoP (CA) consistently serves the most traffic.

If we drop cp3 and do not replace it with an Asia/Oceania cp and invest in another in Canada, we would increase GB's traffic by ~29k requests (63% increase) and increase CA's by ~ 24k requests (34% increase).

I am open to suggestions and opinions from MediaWiki (SRE) as to what people think likely be the best investment proposal by Infrastructure in this regard.

Looking at things it seems to me that another CA cp* would be the best way forward for the reasons above.