Naming

1
Naming

• Sandhya Turaga
• CS-249
• Fall-2005

2
Outline of Chapter

• Naming in general
• Characteristics of distributed naming
• Bindings
• Consistency
• Scaling
• Approaches to Design a global name service
• DEC Global Name Service ( by Lampson)
• Stanford Design (by DAVID R. CHERITON and TIMOTHY
  P. MANN)

3
Naming in general

• Names facilitate sharing.
• Names are used for communication of objects
• Unique identifiers
• Never reused, always bound to an
  object, provides location independence
• Pure Names
• Nothing but a bit pattern that is an
  identifier
• Pure names commit one to nothing
• Eg 12306521
• Impure Names
• Carries commitments of one sort or the
  another
• Eg src/dec/ibm

4
Characteristics

• Bindings
• Machine to address binding
• Services to machine binding
• Consistency
• How one propagates updates among replicates
  of a naming database (Explained more in the DEC
  example)

5
Characteristics ( contd)

• Scaling
• Being able to manage indefinite number of
  machines each providing some part of name lookup
  service of indefinite size which is managed by a
  great number of more or less autonomous
  administrators without any major changes to the
  existing environment.
• Managing name Space rather than the avoidance
  of scale disasters In the implementation

6
Approaches

• DEC Global Naming Service
• Has two levels
• Client Level
• Administration Level
• The Stanford Design
• Has three levels
• Global Level
• Administrational Level
• Managerial Level

7
Approach DEC client level

• Client level
• File structure is seen similar to a UNIX file
  system - like a tree of directories ( Figure 1)
• Each directory is identified by a unique
  Directory Identifier (DI)
• Each ARC of the directory is called a directory
  reference (DR) DI of the child directory
• Directory path relative to the root is Full Name
  (FN)
• In the figure1 finance/personal is relative
  to root src

8
File system as seen in DEC client level (Figure
1)
DR300
DR200
9
DEC client level (contd)

• In this tree structure each node carries
  timestamp a mark which is either absent/present
• A node is present/absent can be known by the
  update operations.
• A node is which is absent is denoted by a strike
  on the circle in figure 2.
• Time-stamps are used to allow trees to be updated
  concurrently.

10
Figure 2
11
DEC-Administrational Level

• Administrator allocates resources to the
  implementation of the service and reconfigures it
  to deal with failures.
• It sees a set of directory copies (DC), each one
  stored on a different server (S) machine.
• Figure 3 shows dec/src is stored on four servers.

12
Figure 3
13
DEC- Administration level (Contd..)

• A lookup can try one or more of the servers to
  find a copy from which to read.
• The values on servers(10,12) are the current time
  which is called the last sweep time.
• Each copy also has nextTS (next time-stamp) it
  will assign to a new update.
• Updates are spread through Sweep operation which
  has a time-stamp sweepTS.
• An update originates at one DC

14
DEC- Administration level (Contd..)

• Sweep operation sets the DCs nextTS to sweepTS
  and then reads collects updates from DC.
• Once these updates are written back the lastsweep
  is set to sweepTS.
• Messages are sent among DCs with updates to speed
  up the update propagation. ( figure 4-a )
• After a successful sweep operation at time 14
  Figure 3 would look like Figure 4-b

15
Figure 4-a
16
Figure 4-b
17
DEC- Administration level (Contd..)

• To obtain the set of DCs reliably ,all the DCs
  are linked in a ring with the arrows pointing to
  other DC (Figure 5).
• Sweep starts at one DC follows the arrows and
  completes the ring .
• In case of a server a failure ring must be
  reformed. In this process updates may lost and a
  new sweep operation must be started.

18
Figure 5
19
DEC-Name Space

• Expanding the name space and changing its
  structure without destroying the usefulness of
  old names is important in hierarchical naming
  structure.
• Each directory in this hierarchical structure is
  a context within which names can be generated
  independently of what is going on in any other
  directory.
• Thus we can have a directory name Finance under
  two prefixes such as src/dec/Finance
  src/ibm/Finance. Each directory can have its own
  administrator, and they do not have to coordinate
  their actions.

20
Figure 6
444
222
21
DEC-Name Space

• Growth of name space in by combining existing
  name services ,each with its own root is
• add a new root by making the existing root
  nodes as its children (Figure 6).

22
Figure 7
23
DEC-Name Space (Contd..)

• How can a directory be found from its DI?
• Root keeps a table of well-known directories that
  maps certain DIs into links which are FNs
  relative to the root.
• See figure 7.

24
DEC-Name Space (Contd..)

• Sometimes restructuring (Moving a subtree) is
  required.
• For example DEC buys IBM then the path for IBM
  changes as shown in figure 8.
• Old paths as ansi/ibm wont work now. Users
  should be forwarded from ansi/ibm to ansi/dec/ibm
  to access IBM.

25
Figure 8
26
DEC-Name Space - Caching

• Caching is a good idea for lookups.
• Caching is achieved by
• 1) Slow rate of change on the naming database
• 2) Or Tolerating some inaccuracy in caching
  data
• Enforcing slow rate of change can be achieved
  with
• expiration time (TX) on entries in the
  database with an exception.

27
DEC-Name Space (Figure 9)
999/DEC 311 Valid until 20 Nov 2006
28
DEC Name Space

• For example in figure 7 the result of looking up
  ansi/dec/src is valid until 20 Nov 2006

29
Stanford Naming System Design

• It has three levels
• Global Level
• Administrational level
• Managerial level
• In lower levels, naming is handled by object
  managers
• How do administrational global levels differ?

30
Stanford Naming System Design (Contd..) (Figure
10)
Fig 10
31
Managerial level

• Each directory is stored by a single object
  manager
• Any kind of object can be named using a directory
  implemented by its object manager. Eg Files
• Object manager stores the absolute name of the
  root of each subtree

32
Managerial level (contd ..)

• For example (in Figure l0), the subtrees rooted
  at the directories edu/stanford/dsg/bin and
  edu/stanford/dsg/1ib are both implemented by DSG
  file server 1, which thus covers all the names
  with prefixes edu/stanford/dsg/bin and
  edu/stanford/dsg/lib.
• Accordingly, file server 1 stores all the files
  and directories under these two subtrees

33
Managerial level (Contd..)

• Object manager implements all operations on the
  names it covers.
• What are the advantages of integrating names with
  object management?
• Managerial directories record every name-object
  binding

34
Managerial level (contd..)

• Requirements to construct a complete naming
  service
• Clients should know to which manager it should
  send the messages
• Clients should know which name is unbound and
  which name is unavailable
• Separate mechanism is needed to implement
  operations on directories above the managerial
  level.

35
Managerial level (contd..)

• Finding manager location can be provided by
  prefix caches multicasting.
• Each client in a naming system maintains a name
  prefix cache.
• Each entry of cache associates a name prefix with
  a directory identifier.
• Directory identifier has two fields (m,s)
• m- Manager identifier
• S-Specific directory identifier

36
Managerial level (contd..)

• Hit
• when a cache search returns a cache entry
  containing the name of the managerial directory
  it is called Hit
• near miss
• When a cache search doesnt return the name
  of the directory but does return something is
  called near miss.

37
Managerial level (contd..)

• If cache search returns a local administrational
  entry then the client multicasts a Probe
  request to a group of managers specified by the
  cache entry.
• If the cache search returns a directory
  identifier that specifies a liaison server then
  the client multicasts a probe request on the
  given name to liaison server.
• What is liaison server?

38
Managerial level (contd..)

• What if the liaison server crashes in between?
• What happens when the liaison server receives the
  probe?
• Cache consistency is maintained by discarding
  stale cache entries.
• What is Stale cache entry?

39
Managerial level (contd..)

• Advantages of name prefix caching mechanism
• 1) High cache-hit ratio Performance.
• 2) Near miss reduces the amount of work
  required for the shared naming system.
• 3) The longer the prefix returned by the
  near miss the more work is saved.
• 4) Correctness of cache information is
  automatically checked.

40
Administrational level

• Administrational directories are implemented
  using object managers and administrational
  directory managers.
• Administrational directory manager covers the
  unbound names
• Bound names are covered by object managers

41
Administrational level (contd..)

• The administrational directory manager holds a
  list of bound names, but it is not considered to
  cover these names
• Figure 11 illustrates how information is
  distributed in the directory edu/stanford/dsg/use
  r of Figure10

42
Figure 11
Fig 11
43
Figure 10(Same as in slide 30)
44
Administrational level (Contd..)

• Managers that cooperate in implementing an
  administrational directory are called its
  participants they form a participant group.
• Each participant in the admin directory responds
  to probes on the names it covers.
• Every name is covered by at least one
  participant.

45
Administrational level (Contd..)

• Directory listing is maintained by directorys
  manager.
• Clients obtain list of names from directory
  manager
• What if directory managers fail?
• Directorys manager can coordinate access to the
  directory by clients located outside the local
  administration

46
Administrational level (Contd..)

• Remote administrational directory can be accessed
  through local liaison server and the global
  directory system.
• Advantages of this technique Even if directory
  manager is down the corresponding file server can
  respond to name-mapping requests
• ( see figure 12)

47
Figure 12
Global level
Administration level
managerial level
48
Global level

• This design is similar to Global directory system
  of DEC implementation proposed by Lampson.
• Interfacing to the global directory system
• Liaison servers act as intermediaries for all
  client operations at the global level.
• Caching performed by liaison servers improves the
  response time for global level queries

49
Global level (Contd..)

• Also reduces load on the global directory system.
• What happens if a global directory system becomes
  unavailable within some administration?

50
Performance of name prefix caching

• Load per operation
• 1) A packet event is transmission or reception
  of network packet.
• 2) Multicast with g recipients costs g1pkt
  events
• 3)The avg number of pkt events required to
  map an event
• Cmap4h (r m 7)(1 - h)

51
Performance of name prefix caching

• h- Cache hit ratio
• r- No.of retransmissions required to know the
  host is down
• m-No.of object managers in the system
• Derivation of the above equation
• when there is cache hit name mapping costs -4
  pkt events
• when there is a cache miss rm7pkt events

52
Performance of name prefix caching

• When can Cmap reach the optimum value?
• What is this rm7?
• Is there a statistical model for computing a
  cache-hit ratio? (See next slide)
• Cache-performance model

53
Cache-performance model

• Input parameters
• 1) No.of name-mapping requests issued per
  unit time
• 2) The average length of time a name-cache
  entry is valid
• 3)The average length of time a client cache
  remains in use before it is discarded
• 4)The locality of reference

54
Cache-performance model

• Avg. steady-state hit ratio of all clients
• h1-? ? ß/( ßj,k vk )
• j k
• ßj,k Avg interarrival time for requests
  generated by client j that reference a name in
  managerial sub tree k.
• Vk Validity time for cache entry of a managerial
  sub tree k.
• ß-Global Avg interarrival time for name-mapping
  requests.

55
Cache-performance model

• Steady-state ratio for single pair is
• hj,k 1- ßj,k /( ßj,k vk )
• Miss-ratiototal no.of misses/total no.of
  requests
• Hit-ratio1-(Miss-ratio)

56
Cache performance model (Contd..) (Figure 13)
Fig 13
57
Cache performance model (Contd..)

• Is it reasonable to expect the ratio of h
• in the range 99.00-99.98?
• The following graph shows h for each
• client, sub tree pair
• Vk varies from 100-5000(on x-axis)
• Hit-ratio is 0.9901 at Vk 100 0.9998 for Vk is
  5000.

58
Cache performance model (Contd..) Figure 14
Fig 14
59
Cache performance model (Contd..)

• What is startup misses ?
• Will startup misses effect hit-ratio ?
• Depends on validity time
• ( see figures 15 )

60
Hit Ratio as a function of Time ( figure 15 )
61
Conclusions

• How large can a system be built with this design
  ?
• Depends on load per operation
• Load on managers
• ( see figure 16 )

62
Load per server as a function of system size -
Figure 16
63
Questions ?