Nap - NUMA-Aware Persistent Indexes

---

Evaluation Environment

We use a 4-socket server (equipped with Optane DC Persistent Memory) to evaluate Nap, which you can log into. Please check https://osdi21ae.usenix.hotcrp.com/ for the access method of our server.

The detailed information of the server:

• 4 * 18-core Intel Xeon Gold 6240M CPUs
• 12 * 128GB Optane DIMMs
• 12 * 32GB DDR4 DIMMs
• Ubuntu 18.04 with Linux kernel version 5.4.0

You can run ipmctl show -dimm and ndctl list to show the configuration of Optane DC PM.

The server has installed dependencies of Nap, including PMDK, tbb and google-perftools.

Before the start

After logging into our server, run cd /home/wq/Nap. This is path of our codebases.

Note:

1. Before evaluating Nap, please run w to check if anyone else is using it, to avoid resource contention. And please close ssh connection when not conducting evalation.
2. If you find the performance is abnormal, please run reboot to reboot this server, since Optane DIMMs may become slower when being writeen continuously.
3. If you have any question, please contact with me via q-wang18@mails.tsinghua.edu.cn

Directory Organization:

• include, src: codes of Nap and PM indexes
• bench: codes of benchmarks
• script: script for AE
• dataset: dataset for evalutaion

Functionality of codes

• include/cn_view.h: GV-View in Section 3.3
• include/sp_view.h: PC-View in Section 3.4
• include/top_k.h, include/count_min_sketch.h: min heap, count-min sketch and logic of hot set identification (Section 3.5)
• include/nap.h: main logic of Nap, function nap_shift is 3-phase switch (Section 3.6)
• include/index/*: PM indexes from https://github.com/chenzhangyu/Clevel-Hashing/ and https://github.com/utsaslab/RECIPE/
• bench/*_nap.cpp: Nap-converted PM indexes.

Main Figures

First, run cd script.

Figure 1

bash ./run_moti_fig1.sh, which prints the bandwith of local/remote read/write, with varying thread counts. About 10~20 mins.

Note: when finishing figure 1, run bash ./setup_eval.sh to initialize ext4-DAX file systems for PMDK, which is used by remaining figures.

Figure 8

execute bash ./run_fig8_{cceh, clevel, clht, masstree, fastfair}.sh, which produces four output files: XX_WI_Raw, XX_RI_Nap, XX_WI_Raw, XX_RI_Nap, where XX is {cceh, clevel, clht, masstree, fastfair}.

Each script needs 90-120 mins to run (fastfair needs > 200mins).

Note: FastFair always triggers segment fault, so you may be run it multiple times.

Note: All output files are saved in /home/wq/Nap/build (i.e., ../build)

Note: If you want to complie the codes manually, please run rm CMakeCache.txt before run cmake ..

Other Figures

If you are interested in reproducing other figures, go ahead :)

Figure 2(b)

This result is the same as Figure 8(a).

Figure 3

bash ./run_fig3.sh, which prints the resulted access ratio.

Figure 9

bash ./run_fig9.sh, producing two output files: Fig9_scan_Raw and Fig9_scan_Nap.

About 60 mins.

Figure 10

bash ./run_fig10.sh, which produces two output files: Fig10_lat_Raw and Fig10_lat_Nap. Each line of these files contains two value: < latency (us), CDF >.

About 10 mins.

Figure 11

We use Intel’s PCM tools to measure the remote PM accesses. The pcm.x sub-tool provides the amount of data through UPI links and the pcm-numa.x sub-tool monitors remote DRAM accesses. Leveraging the two sub-tools, we calculate the remote PM accesses of P-CLHT under write-intensive workloads.

This expriment is time-consuming, since we need to run build/clht_nap multiple times to get stable results.

Figure 12

bash ./run_fig12.sh, which produces two output files: Fig12_3_phase and Fig12_global_lock. Each line of these files contains two value: < time (ms), throughput (ops/ms) >. You need to select a continuous piece of data in the output files to depict figures.

About 10-20 mins.

Figure 13

bash ./run_fig13.sh, which produces two output files: Fig13_NR_WI and Fig13_NR_RI. These files contain the results of NR under write-intensive (WR) and read-intensive (RI) workloads.

About 60 mins.

Figure 14

bash ./run_fig14.sh, which produces 5 output files:

• Fig14_hotset_Nap: Figure 14(a)
• Fig14_keyspace_Nap and Fig14_keyspace_Raw: Figure 14(b)
• Fig14_zipfan_Nap and Fig14_zipfan_Raw: Figure 14(c)

For Figure(d) and (e), since we must add/remove Optane DIMMs manually, we omit their evalation. But if you are interested, you can run bash ./run_fig8_clht.sh on your own machine.

About 40-60 mins.

Table 2

bash ./run_table2.sh, which produces output file Table2_recovery.

Note: FastFair always triggers segment fault, so you may be run it multiple times.

About 20 mins.

Figure 15

We cannot provide environment of this experiment currently, since the client servers equipped with ConnectX-6 NICs are being occupied by others (for conducting experiments for other papers).