# CS 5220 ## Distributed memory ### Modeling message costs ## 06 Oct 2015
### Basic questions - How much does a message cost? - *Latency*: time to get between processors - *Bandwidth*: data transferred per unit time - How does *contention* affect communication? - This is a combined hardware-software question! - Goal: understand just enough to model roughly
### Conventional wisdom - Roughly constant latency (?) - Wormhole routing (or cut-through) flattens latencies vs store-forward at hardware level - Software stack dominates HW latency! - Latencies *not* same between networks (in box vs across) - May also have store-forward at library level - Avoid topology-specific optimization - Want code that runs on next year’s machine, too! - Bundle topology awareness in vendor MPI libraries? - Sometimes specify a *software* topology
### $\alpha$-$\beta$ model Crudest model: $t_{\mathrm{comm}} = \alpha + \beta M$ - $t_{\mathrm{comm}} =$ communication time - $\alpha =$ latency - $\beta =$ inverse bandwidth - $M =$ message size Works pretty well for basic guidance! Typically $\alpha \gg \beta \gg t_{\mathrm{flop}}$. More money on network, lower $\alpha$.
### LogP model Like $\alpha$-$\beta$, but includes CPU time on send/recv: - Latency: the usual - Overhead: CPU time to send/recv - Gap: min time between send/recv - P: number of processors Assumes small messages (gap $\sim$ bw for fixed message size).
### Communication costs Some basic goals: - Prefer larger to smaller messages (avoid latency) - Avoid communication when possible - Great speedup for Monte Carlo and other embarrassingly parallel codes! - Overlap communication with computation - Models tell you how much computation is needed to mask communication costs.
### Intel MPI on Totient - Two 6-core chips per nodes, eight nodes - Heterogeneous network: - Ring between cores - Bus between chips - Gigabit ethernet between nodes - Test ping-pong - Between cores on same chip - Between chips on same node - Between nodes

### Network model

• On-chip: $\alpha$-$\beta$ works well!
• Off-chip: Not so much
• But cross-node communication is clearly expensive!
### Moral Not all links are created equal! - Might handle with mixed paradigm - OpenMP on node, MPI across - Have to worry about thread-safety of MPI calls - Can handle purely within MPI - Can ignore the issue completely? For today, we’ll take the last approach.