Test 2 Study Guide

PDF

Test 2 study guide

NOTE: Some of these answers were produced collaboratively amongst students in the class. I do not claim authorship of every answer.

Summary

• Lamport’s Logical Clock
  • Part A
  • Part B
  • Part C
  • Part D
• Lamport’s ME Lock
• RPC Latency Limits
  • Question 1
    • Part A
    • Part B
  • Question 2
    • Part A
    • Part B
    • Part C
  • Question 2
• Active networks
  • Question 1
    • Part A
    • Part B
  • Question 2
  • Question 3
• SPRING OS
  • Question 1
    • Part A
    • Part B
    • Part C
• Enterprise Java Beans
  • Question 1
    • Part A
    • Part B
  • Question 2
• DSM
  • Question 1
    • Part A
    • Part B
  • Question 2
• GMS
  • Question 1
  • Question 2
  • Question 3
• DFS
  • Question 1
    • Part A
• Potpourri?

Lamport’s Logical Clock

Part A

$E1 \rightarrow E6 \rightarrow E7 \rightarrow E9 \rightarrow E10 \implies E1 \rightarrow E10$

Part B

$E1 \rightarrow E2 \land E1 \rightarrow E8 \implies E2 \parallel E8$

Part C

$E8 \rightarrow E3 \land E3 \rightarrow E4 \land E11 \rightarrow E4 \implies E8 \parallel E11$

Part D

$E2 \rightarrow E4 \land E11 \rightarrow E4 \implies E2 \parallel E11$

Lamport’s ME Lock

• Every process can send to every other process
• Messages can’t be lost
• For any two processes, messages are received in the same order they are sent
• Queues are totally ordered by Lamport’s logical clocks

RPC Latency Limits

Question 1

Part A

T1 + C2 + T2 + C4

Part B

Yes, T1 + C2 + T2, there this is justified if (T1 + C2 + T2) <= (C1 + C4)

This represents the client sleeping/spinning during the RPC call rather than switching to another process and being justified if the latency is less than or equal to the time it takes the two context switches at the client.

Question 2

Part A

Calls from the client acknowledge the previous response from the server. If the server receives an acknowledgment that it has already received, the call must be a duplicate.

Part B

If the server doesn’t buffer the response and the response packet is lost, then you have to re-execute the procedure, which can be more latency intensive than buffering.

Part C

The server will maintain a buffered response until it has a new request from the original client.

Question 2

Downloading the client stub into the kernel requires that the client stub code be trusted, as it will be running inside the protection domain of the kernel.

Active networks

Question 1

Part A

Size of soft cache is limited and as new network flows hit the node, type fields for old flows get evicted.

Part B

This will be an uncommon occurrence, as most flows are composed of many packets in short succession, during which the cache will container the code required for the flow, but several concurrent flows can overwhelm the cache’s capacity.

Question 2

Dropping packets should result in a retransmission which will reload the type into the PREV’s cache.

Question 3

• SDN does not require arbitrary code execution
• SDN is used to setup routing tables which can then run at hardware speeds and not slow down the network

SPRING OS

Question 1

Part A

• A singleton contract between client and load balancer
• A replicated contract between load balancer and webservers

Part B

Specify which clients have the write access in the DB’s Access Control List (ACL). Before a client makes an invocation it will encounter a “Front Object” of the DB domain which will consult the ACL and enforce the policy.

Part C

He/she can create a new client for the file server. The IDL for the file server would already been written, and client side stub code for the file server would already have been generated. We can use the client side stub code to access and invoke the file server object.

Existing web server would continue to function, while we can test the new custom communication by creating a new client for the file server and invoking it. A new subcontract for communication between this new client and file server would be created.

Enterprise Java Beans

Question 1

Part A

The entity beans allow for concurrency within the client call and the servlets allow for concurrency between clients.

Part B

THe business logic is exposed to the clients of the app.

Question 2

• Web container with servlets that correspond to the session of a client with the presentation logic.
• EJB Container with the business logic and entity beans which are the DAOs serving as an interface for the business logic layer to access the DB
• DB Server

This hides business logic, allows concurrency between clients and is reasonably performant.

DSM

Question 1

Part A

As part of the multiple writer page consistency the DSM maintains page diffs for each lock. P1, P2, and P3 are using different locks, so there will be 3 diffs for page X corresponding to each lock. If in the future, some process access X using one of these locks, then updated page X will be determined by applying the diffs of that lock on the updated copy returned by the owner of X.

Part B

• Lock Acquisition
  • Nothing apart from bringing the current copy of X if there is a page fault
• Critical Section
  • A copy of the page X will be created and the original page X will be left to be modified by the process
• Lock Release
  • A diff will be calculated between the modified X and its copy. The copy will be deleted and the diff will be stored as the metadata for X

Question 2

• Lock Acquisition
  • As page X has been modified earlier by P2 and P5, pages held by the previous lock owner will be invalidated.
• Critical Section
  • There will be a page fault as P6 tries to access X. DSM will bring the current valid X from its owner and will apply the diffs. A copy of X will be made.
• Lock Release
  • A new diff will be calculated using the current X and the copy, and the copy will be deleted.

GMS

Question 1

• Node P reads page X from the disk and puts it in its local page cache
• Node R evicts page Y from its memory and writes it to disk if its dirty
• If P has pages in its global cache, one of the global pages is moved from P to R
• If P does not have any pages in its global cache, the LRU page on P is moved from P to R. If the LRU local page is dirty, then it needs to be written to disk as global pages are all considered clean.
• If P and R happen to be the same node, then the algorithm would be the same except there would be no transfer of pages since P had the globally oldest page.

Question 2

• Master node recomputes the partitions of the UID space for each node (including new node)
• Each node sends the GCD entries that it is no longer responsible for to the new node

The stripe group and MMap are also replicated globally so it would need to be adjusted and added to N

Question 3

• The only “global” action is at the beginning of every epoch when the weight vectors and minAge has to be computed by the initiator.
• During the epoch itself, all actions at a node are entirely based on local decision-making.

DFS

Question 1

Part A

his approach fails to meet conditions bullet 3. If we keep the metadata manager static then if that file becomes popular that metadata manager will be swamped with requests (creating a hotspot)

Potpourri?

JAVA RMI???