Tardis 2.0 Eval

tardis-2.0-eval

Tardis 2.0: Optimized Time Traveling Coherence for Relaxed Consistency Models — Yu and Devadas (2015) (arXiv:1511.08774, 2015)

What this evaluates

Evaluates the performance, network traffic, and hardware overhead of the Tardis 2.0 cache coherence protocol under Total Store Order (TSO) consistency. It measures speedup and traffic reduction against a full-map MSI directory baseline and a baseline Tardis implementation across 20 multicore system benchmarks.

Datasets

• Splash2, PARSEC, HPCG, YCSB, TPCC — total 20; splits: test (20)

Metrics

• speedup (primary) — range: other
  • Ratio of baseline execution time to Tardis 2.0 execution time. All numbers are normalized to the full-map MSI directory protocol running under Sequential Consistency (SC).
• renew rate — range: percent
  • Ratio of the number of renew requests to the total number of Last-Level Cache (LLC) accesses.
• network traffic — range: other
  • Total volume of inter-core and memory traffic, broken down into DRAM, common, renew, and invalidation components.

Input / output format

Input: Simulator configuration parameters (core count, cache sizes, network topology) and benchmark workloads from Splash2, PARSEC, HPCG, YCSB, and TPCC.

Output: Execution traces, cycle counts, cache coherence message logs, and aggregated performance/traffic metrics normalized to the directory SC baseline.

Scoring recipe

def calculate_metrics(baseline_time, tardis_time, baseline_traffic, tardis_traffic, renew_req, llc_acc):
    speedup = baseline_time / tardis_time
    traffic_reduction = (baseline_traffic - tardis_traffic) / baseline_traffic
    renew_rate = renew_req / llc_acc
    return speedup, traffic_reduction, renew_rate

Common pitfalls

• Speedup values are normalized to the directory protocol under SC, not TSO, so direct comparison to TSO baselines requires adjusting the reference point.
• Renew rate is specific to Tardis's lease-based coherence and does not apply to directory-based protocols, making cross-protocol traffic comparisons asymmetric.
• Hardware overhead is reported per-core but scales with the number of cache lines in L1/LLC, not just core count.

Evidence (verbatim from paper)

Our experiments are executed over 20 benchmarks selected from Splash2 [19], PARSEC [20], sparse linear algebra [21] and OLTP database applications [22]. Fig. 3 shows the speedup of directory coherence and Tardis running SC and TSO. All numbers are normalized to the directory protocol with SC. We define renew rate as the ratio of the number of renew requests over the total number of LLC accesses.

Citation

@misc{yu2015tardis2,
  title={Tardis 2.0: Optimized Time Traveling Coherence for Relaxed Consistency Models},
  author={Yu and Devadas (2015)},
  year={2015},
  note={arXiv:1511.08774}
}

• arXiv: 1511.08774