---

Cost simplication in distributed data processing

---

• Communication cost in distributed data processing
  • Fact:

    Data transmission rates over long haul network links is relatively low due to: Communication protocol overhead Reliable communication requires that each packet is acknowledged (by an ACK messages) Packet queueing delay --- especially when the network is congested !!!           There are many routers between the source node and the destination node In each router, a packet must wait for its turn to be transmitted on the output link !!!

    ---

    ---

  • Example:

    Time to access 1 data block (~ 4 KBytes) from disk: approximately: 10 msec Raw transmission time to transfer 4 KBytes of data with a 1 Mbps long haul link: 4 KBytes = 4000 × 8 = 32,000 bits 32,000 Time to transmit 32,000 bits = ----------- × 1000 msec 1,000,000 = 32 msec Real transmission time to transfer 4 KBytes of data with a 1 Mbps long haul link: cheung@home6(53)> ping -s compute PING compute: 56 data bytes 64 bytes from compute.mathcs.emory.edu: icmp_seq=0. time=46.2 ms 64 bytes from compute.mathcs.emory.edu: icmp_seq=1. time=58.6 ms 64 bytes from compute.mathcs.emory.edu: icmp_seq=2. time=49.6 ms 64 bytes from compute.mathcs.emory.edu: icmp_seq=3. time=61.7 ms Transmission delay from my home to Emory = 50 msec (from ping) Transmit 4 KBytes = 32 msec Transmission delay for ACK packet = 50 msec ============================================================ Total time to transmit a 4 KBytes packet = 132 msec !!!

  ---

  ---

  ---

  ---

• Simplication in cost computation of distributed data processing
  • Simplifying assumption:

    Data communication is more costly than disk accesses in a wide area network (Mbps)

    Common practice in distributed data processing:

    We ignore the disk access time in the performance analysis of Relational Algebra operations (and only consider the communication cost)

  ---

  ---

  ---

  ---

• Example distributed query processing
  • Scenario:

    Site 1 contains the relation Employee Site 2 contains the relation Department Site 3 performs the query    πfname,lname,dname ( Employee ⋈dno=dnum Department )

    ---

    ---

  • Cost of computing π_{fname,lname,dname} ( Employee ⋈_dno=dnum Department ) :

    Option 1: Communication cost = 1,003,500 bytes Option 2: Communication cost = 1,400,000 bytes Option 3: Communication cost = 403,500 bytes