20-30% Performance Hit from the Spectre Bug Fix on Ubuntu
In this blog post, we’ll look at the performance hit from the Spectre bug fix on Ubuntu.
Recently we measured the performance penalty from the Meltdown fix on Ubuntu servers. It turned out to be negligible.
Today, Ubuntu made a Spectre bug fix on Ubuntu available, shipped in kernel 4.4.0-112. As with the Meltdown fix, we measured the effect of this update. Unfortunately, we observed a major performance penalty on MySQL workloads with this new kernel.
Our benchmark used the following:
System:
- CPU:
- 2 x Intel(R) Xeon(R) CPU E5-2680 v3 @ 2.50GHz (Codename Haswell)
- /proc/cpuinfo has 48 entries
- Server
- Supermicro Motherboard X10DRI
- Storage
- NVMi Intel DC 36000
- Ubuntu 16.04
- no-fix – kernel 4.4.0-104-generic
- fix (now GA) – kernel 4.4.0-112-generic
With the fix, the Spectre and Meltdown mitigation detection tool v0.27 reports following:
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 |
Checking for vulnerabilities against live running kernel Linux 4.4.0-112-generic #135-Ubuntu SMP Fri Jan 19 11:48:36 UTC 2018 x86_64 CVE-2017-5753 [bounds check bypass] aka 'Spectre Variant 1' * Checking count of LFENCE opcodes in kernel: YES > STATUS: NOT VULNERABLE (115 opcodes found, which is >= 70, heuristic to be improved when official patches become available) CVE-2017-5715 [branch target injection] aka 'Spectre Variant 2' * Mitigation 1 * Hardware (CPU microcode) support for mitigation: YES * Kernel support for IBRS: YES * IBRS enabled for Kernel space: YES * IBRS enabled for User space: NO * Mitigation 2 * Kernel compiled with retpoline option: NO * Kernel compiled with a retpoline-aware compiler: NO > STATUS: NOT VULNERABLE (IBRS mitigates the vulnerability) CVE-2017-5754 [rogue data cache load] aka 'Meltdown' aka 'Variant 3' * Kernel supports Page Table Isolation (PTI): YES * PTI enabled and active: YES > STATUS: NOT VULNERABLE (PTI mitigates the vulnerability) |
Workload:
- sysbench 1.0.11 oltp_read_only and oltp_point_select
- 32 tables, 10mln rows each
- datasize ~73GB
Server:
- Percona-Server-5.7.20-19-Linux.x86_64 binary distribution
- config https://github.com/Percona-Lab-results/201701-meltdown-OLTP-ro/blob/master/my.cnf
Sysbench script:
The results
The results are in transactions per sec (more is better).
In memory (buffer pool 100G)
Connection via TCP
| bp | workload | threads | tps no-fix | tps fix | ratio no-fix/fix |
|---|---|---|---|---|---|
| 100 | oltp_read_only | 1 | 810.76 | 655.34 | 1.24 |
| 100 | oltp_read_only | 2 | 1589.66 | 1277.02 | 1.24 |
| 100 | oltp_read_only | 8 | 6233.69 | 5018.11 | 1.24 |
| 100 | oltp_read_only | 16 | 11253.28 | 9477.18 | 1.19 |
| 100 | oltp_read_only | 64 | 22702.29 | 18564.30 | 1.22 |
| 100 | oltp_read_only | 128 | 22281.22 | 18357.06 | 1.21 |
| 100 | oltp_point_select | 1 | 16095.32 | 12380.20 | 1.30 |
| 100 | oltp_point_select | 2 | 32665.97 | 24907.84 | 1.31 |
| 100 | oltp_point_select | 8 | 132480.34 | 101787.44 | 1.30 |
| 100 | oltp_point_select | 16 | 236832.94 | 189087.10 | 1.25 |
| 100 | oltp_point_select | 64 | 498322.41 | 415631.91 | 1.20 |
| 100 | oltp_point_select | 128 | 496661.65 | 414495.64 | 1.20 |
Buffer pool 50G
Connection via TCP
| bp | workload | threads | tps no-fix | tps fix | ratio no-fix/fix |
|---|---|---|---|---|---|
| 50 | oltp_read_only | 1 | 683.09 | 595.63 | 1.15 |
| 50 | oltp_read_only | 2 | 1390.70 | 1143.30 | 1.22 |
| 50 | oltp_read_only | 8 | 5262.02 | 4493.87 | 1.17 |
| 50 | oltp_read_only | 16 | 9842.04 | 8242.02 | 1.19 |
| 50 | oltp_read_only | 64 | 21021.20 | 17644.76 | 1.19 |
| 50 | oltp_read_only | 128 | 21526.21 | 17932.34 | 1.20 |
| 50 | oltp_point_select | 1 | 14535.73 | 11758.57 | 1.24 |
| 50 | oltp_point_select | 2 | 28721.43 | 23277.60 | 1.23 |
| 50 | oltp_point_select | 8 | 108422.96 | 90189.94 | 1.20 |
| 50 | oltp_point_select | 16 | 203876.31 | 167382.92 | 1.22 |
| 50 | oltp_point_select | 64 | 447757.48 | 376506.97 | 1.19 |
| 50 | oltp_point_select | 128 | 473894.73 | 384301.33 | 1.23 |
Buffer pool 25G
connection via TCP
| bp | workload | threads | tps no-fix | tps fix | ratio no-fix/fix |
|---|---|---|---|---|---|
| 25 | oltp_read_only | 1 | 542.09 | 470.88 | 1.15 |
| 25 | oltp_read_only | 2 | 1074.54 | 931.02 | 1.15 |
| 25 | oltp_read_only | 8 | 4169.10 | 3621.79 | 1.15 |
| 25 | oltp_read_only | 16 | 7626.30 | 6716.29 | 1.14 |
| 25 | oltp_read_only | 64 | 18206.18 | 15702.90 | 1.16 |
| 25 | oltp_read_only | 128 | 20224.22 | 16966.13 | 1.19 |
| 25 | oltp_point_select | 1 | 11107.73 | 9294.73 | 1.20 |
| 25 | oltp_point_select | 2 | 22486.65 | 18526.84 | 1.21 |
| 25 | oltp_point_select | 8 | 86385.70 | 73226.44 | 1.18 |
| 25 | oltp_point_select | 16 | 161409.65 | 135689.48 | 1.19 |
| 25 | oltp_point_select | 64 | 370809.49 | 320848.79 | 1.16 |
| 25 | oltp_point_select | 128 | 433324.54 | 358947.61 | 1.21 |
We can see that in CPU-bound workloads the overhead is 20-25%, reaching up to 30% in point select queries. In IO-bound (25G buffer pool) workloads, the observed overhead is 15-20%.
This is a major performance hit, and you should consider it carefully before upgrading to the new kernel.
One hope that is retpoline kernels will have much less impact.
Vadim – thanks for sharing these results. Do you have any AMD-based server hardware to see if AMD CPUs are also affected as much as Intel CPUs?
John,
I do not have modern AMD servers on my hands right now
The performance hit relates to the generation of the Intel CPU.
In this case, the title should be that the performance hit relates to E5-2680 CPU (released in 2012, EOL since 2015).
The used CPU is E5-2680 v3, notice the V3 at the end.
Launched in Q3’14 and is not EOL.
Non the less, I’d like some Skylake benchmarks. There are stories that Skylake and newer generations have a smaller performance hit.
I don’t know if it’s related, but I have found a similar performance hit on AMD Ryzen with Sybase SQLAnywhere Database. Doing a large insert from a select was about 20% slower on patched Ubuntu 16.04 compared with unpatched 14.04. However, the really strange result was that execution of a dbbackup command (which runs an ancient single-threaded Java app to capture pages to a backup file) was about 10 times slower! I suspect that says more about the way in which that code was written than is generally useful.