NVIDIA GTC 2018: Spectre/Meltdown Impact on High Performance Workloads
4 likes520 views
This document summarizes a presentation given by Jeremy Eder from Red Hat Performance Engineering about the impact of the Spectre and Meltdown vulnerabilities on high performance workloads. It discusses performance data showing the impact on tasks like TensorFlow training on GPUs. It also provides suggestions for managing the performance impact, such as enabling transparent huge pages and PCID, and disabling specific kernel patches if needed. A list of additional reference links is also included.
NVIDIA GTC 2018: Spectre/Meltdown Impact on High Performance Workloads
- 1. JEREMY EDER - RED HAT PERFORMANCE ENGINEERING1 Spectre/Meltdown Impact on High Performance Workloads Red Hat Performance Engineering NVIDIA GTC, March 2018
- 2. JEREMY EDER - RED HAT PERFORMANCE ENGINEERING2 ● Context ● Performance Data and Capacity Planning ● KBase Overview ● Going Forward AGENDA
- 3. JEREMY EDER - RED HAT PERFORMANCE ENGINEERING3 VULNERABILITY OVERVIEW Meltdown - Rogue data cache load ● Allows reading all memory Spectre Variant 1 - Bounds check bypass ● Throw-away data left insecure in CPU cache Spectre Variant 2 - Branch Target Injection ● CPU branch predictor trained by exploit
- 4. JEREMY EDER - RED HAT PERFORMANCE ENGINEERING DETERMINING VULNERABILITY • OS vendors provide tools to determine vulnerability • Recent Linux and RHEL kernels include the following new sysfs entries 4 $ ls -1 /sys/devices/system/cpu/vulnerabilities/* /sys/devices/system/cpu/vulnerabilities/meltdown /sys/devices/system/cpu/vulnerabilities/spectre_v 1 /sys/devices/system/cpu/vulnerabilities/spectre_v 2
- 5. JEREMY EDER - RED HAT PERFORMANCE ENGINEERING Userspace vs. Kernelspace Userspace ( e.g. /bin/bash) Operating System (e.g. Linux kernel) System Call Interface 5
- 6. JEREMY EDER - RED HAT PERFORMANCE ENGINEERING6 PERFORMANCE DATA AND CAPACITY PLANNING
- 7. JEREMY EDER - RED HAT PERFORMANCE ENGINEERING7 TensorFlow on NVIDIA GPUs, CUDA 9.1 %diff vs Pre-CVE Kernel Baseline
- 8. JEREMY EDER - RED HAT PERFORMANCE ENGINEERING8 KERNEL BYPASS NETWORKING - Broadwell
- 9. JEREMY EDER - RED HAT PERFORMANCE ENGINEERING9 CAPACITY PLANNING
- 10. JEREMY EDER - RED HAT PERFORMANCE ENGINEERING10 ● RHEL 6/7 have transparent hugepages enabled by default ○ Reduces amount of TLB entries and thus total flush impact ● RHEL 7.4.z has PCID support from 4.14 ○ Reduces impact of TLB flushes by tagging/tracking ○ (2010+ CPUs have it...check /proc/cpuinfo) ● RHEL has runtime knobs to disable patches (no reboot req’d) HOW TO MANAGE PERFORMANCE IMPACT echo 0 > /sys/kernel/debug/x86/pti_enabled echo 0 > /sys/kernel/debug/x86/ibrs_enabled echo 0 > /sys/kernel/debug/x86/retp_enabled
- 11. JEREMY EDER - RED HAT PERFORMANCE ENGINEERING11 IMPACT VARIES BY WORKLOAD
- 12. JEREMY EDER - RED HAT PERFORMANCE ENGINEERING12 ● Community landing page https://meltdownattack.com/ ● Meltdown and Spectre in 3 Minutes ● FOSDEM Closing Keynote ● Speculative Execution Exploit Performance Impacts ● Controlling the Performance Impact of Microcode and Security Patches ● Intel.com: Advancing Security at the Silicon Level LINKS
- 13. JEREMY EDER - RED HAT PERFORMANCE ENGINEERING THANK YOU plus.google.com/+RedHat linkedin.com/company/red-hat youtube.com/user/RedHatVideos facebook.com/redhatinc twitter.com/RedHatNews