YCSB

• The original YCSB is a key-value store benchmark [4] that does not specify transactions. In our evaluation, we group 10 unique key accesses (either read or write) in a single transaction. We use a total of 10M keys, and each row is 1000 bytes.
• Main functions goes to yscb_workload.cc
• some ycsb conf

    if (Options::kYcsbContentionKey) {
      ycsb::Client::g_contention_key = Options::kYcsbContentionKey.ToInt();
    }
    if (Options::kYcsbSkewFactor) {
      ycsb::Client::g_theta = 0.01 * Options::kYcsbSkewFactor.ToInt();
    }
    if (Options::kYcsbReadOnly)
      ycsb::Client::g_extra_read = Options::kYcsbReadOnly.ToInt();

    ycsb::Client::g_dependency = Options::kYcsbDependency;

• start one loader and one ycsb client

remote YCSB

Design 1

• Future value for each txn
• RMWTxn::WriteRow(TxnRow vhandle) wait for some value from the remote node to read, write those value to the first 100 bytes of the vhandle row

void RMWTxn::WriteRow(TxnRow vhandle)
{
  auto dbv = vhandle.Read<Ycsb::Value>();
# receive future value
  dbv.v.assign(Client::zero_data, 100);
  dbv.v.resize_junk(999);
  vhandle.Write(dbv);
}

• RMWTxn::ReadRow(TxnRow vhandle) if it is a remote txn, send the values read from current vhandle to the remote nodes

void RMWTxn::ReadRow(TxnRow vhandle)
{
  vhandle.Read<Ycsb::Value>();
# send future value
}

• need to have bitmap to identify which AttachRoutine is remote
• Reuse the design old future value memory, the ycsb will be perform same as tpcc
• The problem is that it still have problem as tpcc we had before, how should this be different from the old payment remote txn

Design 2

• force some ycsb txn to read from remote txn
• should we as for the value and let remote nodes sends back? the previous simpilfy the design by only let the remote nodes send back
• From Hermes paper:

A local YCSB transaction models a user who puts two products from the same category into his shopping cart. On the other hand, a distributed YCSB transaction models a user who puts a product from a category and another product from all possible categories (which can be modeled by a global Zipfian distribution) to his cart.

• this means we need to classify local ycsb tables into one category, and tables in the another node into another category, all category needs to be known beforehand like how we treat warehouse_id and how warehouse_id seperates on different nodes

• Design in a way, read txn can be remoted compared to previous design that remote read is served for local write.