PERFSONAR LS SCALABILITY ISSUES 1. Common LS queries The most common LS query will likely be: –...

7
PERFSONAR LS SCALABILITY ISSUES 1

description

Sample Performance Results Find the topology server for ESnet –Queries only 1 hLS –Query time = 14 seconds Find all pSB MAs with throughput data for host albu-pt1.es.net –Queries a total of 12 hLS –Total Query time = 110 seconds Find all pSB MAs for community “LHC” –Queries a total of 26 hLS –Query time = 179 seconds 3

Transcript of PERFSONAR LS SCALABILITY ISSUES 1. Common LS queries The most common LS query will likely be: –...

Page 1: PERFSONAR LS SCALABILITY ISSUES 1. Common LS queries The most common LS query will likely be: – “find me the service_accesspoint for X” Samples: –Find.

1

PERFSONAR LS SCALABILITY ISSUES

Page 2: PERFSONAR LS SCALABILITY ISSUES 1. Common LS queries The most common LS query will likely be: – “find me the service_accesspoint for X” Samples: –Find.

2

Common LS queries

• The most common LS query will likely be:– “find me the service_accesspoint for X”

• Samples:– Find me the topology server for ESnet– Find me all pSB MAs for community “LHC”– Find me all pSB MAs with throughput data for host X

• Currently the last 2 queries all need to hit many hLSes.– This fundamentally can not scale!– Issues: Some hLS will be down, when to timeout?,

very high latency to hLS on the other side of the planet, etc.

Page 3: PERFSONAR LS SCALABILITY ISSUES 1. Common LS queries The most common LS query will likely be: – “find me the service_accesspoint for X” Samples: –Find.

3

Sample Performance Results

• Find the topology server for ESnet– Queries only 1 hLS – Query time = 14 seconds

• Find all pSB MAs with throughput data for host albu-pt1.es.net– Queries a total of 12 hLS – Total Query time = 110 seconds

• Find all pSB MAs for community “LHC”– Queries a total of 26 hLS – Query time = 179 seconds

Page 4: PERFSONAR LS SCALABILITY ISSUES 1. Common LS queries The most common LS query will likely be: – “find me the service_accesspoint for X” Samples: –Find.

4

Solution?

• More data needs to be in (or “next to”) the gLS– See Martin’s proposed solution

• I think it should be easy to answer queries like this using the gLS only:– Find all pSB MAs for community “LHC”

• Other queries will be trickier, but should still be possible? – Find all pSB MAs with throughput data for host albu-

pt1.es.net

Page 5: PERFSONAR LS SCALABILITY ISSUES 1. Common LS queries The most common LS query will likely be: – “find me the service_accesspoint for X” Samples: –Find.

Fixing things

• gLS architecture is fine and works for summarization and scalability

• The problem is that the gLS doesn’t have service access points, which necessitates contacting each hLS in the current implementation

• Putting SAPs in the gLS would reduce available summarization

5

Page 6: PERFSONAR LS SCALABILITY ISSUES 1. Common LS queries The most common LS query will likely be: – “find me the service_accesspoint for X” Samples: –Find.

Solution Direction

• The original model included other layers of LSes

• Automatic creation of hierarchy is difficult– that’s basically why it was dropped in the

evolution of the dLS architecture• We need a way to discover services

without touching all hLSes– especially by each node

6

Page 7: PERFSONAR LS SCALABILITY ISSUES 1. Common LS queries The most common LS query will likely be: – “find me the service_accesspoint for X” Samples: –Find.

Solution Space

• Two solutions:• Index Services

– Gather relevant information (servicing a particular query) and keep it up to date

– Ex: All pSB MPs, all North American MPs– This would talk to all hLSs and refresh fairly

periodically (pull model)• Domain Lookup Services

– Large domains host DomLS

7