Bug 19250

Summary:	Sunspider gets progressively slower
Product:	WebKit	Reporter:	Gavin Sherlock <gsherloc>
Component:	JavaScriptCore	Assignee:	Nobody <webkit-unassigned>
Status:	RESOLVED WORKSFORME
Severity:	Normal	CC:	ap, barraclough, prashantnevase, puurtuur
Priority:	P2
Version:	528+ (Nightly build)
Hardware:	Mac
OS:	OS X 10.5
URL:	http://webkit.org/perf/sunspider-0.9/sunspider.html

Gavin Sherlock

Reported 2008-05-25 22:57:23 PDT

Steps to reproduce: 1. Launch webkit 2. Go to bug URL 3. Run Sunspider - make note of results, and results' URL 4. Re-run it several times Expected results: Performance should be similar across runs Actual results: Performance is worse after each run - the overall time was always worse. Results of a comparison between the first run and the fifth run pasted below: TEST COMPARISON FROM TO DETAILS ============================================================================= ** TOTAL **: *1.05x as slow* 1836.8ms +/- 0.8% 1927.0ms +/- 0.3% significant ============================================================================= 3d: *1.07x as slow* 243.2ms +/- 2.1% 259.6ms +/- 0.7% significant cube: *1.13x as slow* 77.4ms +/- 4.2% 87.6ms +/- 0.8% significant morph: ?? 80.4ms +/- 2.6% 81.0ms +/- 1.1% not conclusive: might be *1.01x as slow* raytrace: *1.07x as slow* 85.4ms +/- 0.8% 91.0ms +/- 1.7% significant access: *1.08x as slow* 293.0ms +/- 1.7% 316.6ms +/- 0.8% significant binary-trees: *1.27x as slow* 32.0ms +/- 2.7% 40.8ms +/- 4.0% significant fannkuch: - 112.2ms +/- 2.9% 111.0ms +/- 0.0% nbody: *1.16x as slow* 117.2ms +/- 0.5% 135.6ms +/- 0.5% significant nsieve: - 31.6ms +/- 8.6% 29.2ms +/- 1.9% bitops: - 197.0ms +/- 1.3% 196.8ms +/- 0.9% 3bit-bits-in-byte: - 31.2ms +/- 1.8% 31.0ms +/- 2.8% bits-in-byte: - 37.4ms +/- 1.8% 37.2ms +/- 1.5% bitwise-and: - 55.8ms +/- 1.9% 55.8ms +/- 1.9% nsieve-bits: ?? 72.6ms +/- 3.1% 72.8ms +/- 0.8% not conclusive: might be *1.00x as slow* controlflow: - 21.2ms +/- 2.6% 21.2ms +/- 2.6% recursive: - 21.2ms +/- 2.6% 21.2ms +/- 2.6% crypto: *1.04x as slow* 120.2ms +/- 0.5% 125.4ms +/- 1.1% significant aes: ?? 48.4ms +/- 1.4% 48.6ms +/- 2.3% not conclusive: might be *1.00x as slow* md5: *1.04x as slow* 35.8ms +/- 1.6% 37.2ms +/- 1.5% significant sha1: *1.10x as slow* 36.0ms +/- 0.0% 39.6ms +/- 1.7% significant date: *1.03x as slow* 162.2ms +/- 0.8% 167.6ms +/- 0.4% significant format-tofte: *1.03x as slow* 99.6ms +/- 0.7% 102.6ms +/- 0.7% significant format-xparb: *1.04x as slow* 62.6ms +/- 1.1% 65.0ms +/- 1.4% significant math: *1.12x as slow* 190.0ms +/- 0.7% 212.4ms +/- 0.9% significant cordic: *1.14x as slow* 63.4ms +/- 1.8% 72.0ms +/- 2.7% significant partial-sums: *1.10x as slow* 86.6ms +/- 1.3% 95.6ms +/- 0.7% significant spectral-norm: *1.12x as slow* 40.0ms +/- 0.0% 44.8ms +/- 2.3% significant regexp: *1.01x as slow* 175.6ms +/- 0.6% 177.2ms +/- 0.9% significant dna: *1.01x as slow* 175.6ms +/- 0.6% 177.2ms +/- 0.9% significant string: *1.04x as slow* 434.4ms +/- 0.9% 450.2ms +/- 1.0% significant base64: *1.03x as slow* 70.4ms +/- 1.0% 72.4ms +/- 0.9% significant fasta: ?? 80.8ms +/- 1.3% 81.4ms +/- 1.4% not conclusive: might be *1.01x as slow* tagcloud: *1.04x as slow* 102.8ms +/- 3.1% 106.8ms +/- 1.5% significant unpack-code: *1.02x as slow* 107.6ms +/- 0.6% 110.2ms +/- 0.9% significant validate-input: *1.09x as slow* 72.8ms +/- 1.4% 79.4ms +/- 3.6% significant tested with r34130 on 10.5.2 on a MacPro

Attachments
Add attachment proposed patch, testcase, etc.

Gavin Sherlock

Comment 1 2008-05-27 06:21:50 PDT

Note, this behavior is also observed under 3.1.1 (5525.18). The slow down as a percentage and in terms of milliseconds is not as extreme, though I don't know whether that qualifies this as a regression. A comparison of the 1st and 5th runs under 3.1.1, on the same machine as I used for r34130, is below: TEST COMPARISON FROM TO DETAILS ============================================================================= ** TOTAL **: *1.02x as slow* 2858.6ms +/- 0.3% 2923.8ms +/- 0.3% significant ============================================================================= 3d: *1.04x as slow* 350.4ms +/- 0.8% 364.2ms +/- 0.8% significant cube: *1.09x as slow* 115.8ms +/- 2.1% 125.8ms +/- 0.8% significant morph: - 120.4ms +/- 1.6% 119.2ms +/- 1.4% raytrace: *1.04x as slow* 114.2ms +/- 0.9% 119.2ms +/- 0.9% significant access: *1.03x as slow* 444.4ms +/- 0.7% 456.8ms +/- 0.2% significant binary-trees: *1.03x as slow* 60.4ms +/- 2.3% 62.4ms +/- 1.1% significant fannkuch: ?? 201.6ms +/- 1.5% 203.8ms +/- 0.7% not conclusive: might be *1.01x as slow* nbody: *1.07x as slow* 129.4ms +/- 1.1% 138.2ms +/- 0.4% significant nsieve: - 53.0ms +/- 4.1% 52.4ms +/- 1.3% bitops: ?? 388.2ms +/- 1.2% 388.8ms +/- 0.7% not conclusive: might be *1.00x as slow* 3bit-bits-in-byte: - 57.6ms +/- 1.2% 57.0ms +/- 2.2% bits-in-byte: ?? 85.4ms +/- 2.4% 85.6ms +/- 1.3% not conclusive: might be *1.00x as slow* bitwise-and: - 146.4ms +/- 1.5% 145.8ms +/- 2.1% nsieve-bits: *1.02x as slow* 98.8ms +/- 1.1% 100.4ms +/- 0.7% significant controlflow: ?? 75.2ms +/- 0.7% 75.6ms +/- 0.9% not conclusive: might be *1.01x as slow* recursive: ?? 75.2ms +/- 0.7% 75.6ms +/- 0.9% not conclusive: might be *1.01x as slow* crypto: *1.02x as slow* 202.0ms +/- 0.4% 205.8ms +/- 0.5% significant aes: *1.02x as slow* 66.8ms +/- 0.8% 68.0ms +/- 0.0% significant md5: *1.02x as slow* 67.4ms +/- 1.0% 69.0ms +/- 0.0% significant sha1: *1.01x as slow* 67.8ms +/- 0.8% 68.8ms +/- 1.5% significant date: *1.02x as slow* 254.8ms +/- 0.2% 259.0ms +/- 0.5% significant format-tofte: *1.01x as slow* 116.8ms +/- 0.5% 118.2ms +/- 0.9% significant format-xparb: *1.02x as slow* 138.0ms +/- 0.0% 140.8ms +/- 0.4% significant math: *1.04x as slow* 391.4ms +/- 1.1% 407.4ms +/- 0.6% significant cordic: *1.04x as slow* 152.0ms +/- 1.6% 157.4ms +/- 0.7% significant partial-sums: *1.04x as slow* 167.0ms +/- 1.1% 173.4ms +/- 0.8% significant spectral-norm: *1.06x as slow* 72.4ms +/- 0.9% 76.6ms +/- 0.9% significant regexp: - 185.0ms +/- 0.8% 184.2ms +/- 0.6% dna: - 185.0ms +/- 0.8% 184.2ms +/- 0.6% string: *1.03x as slow* 567.2ms +/- 0.2% 582.0ms +/- 0.6% significant base64: *1.04x as slow* 84.6ms +/- 1.7% 87.8ms +/- 1.2% significant fasta: *1.01x as slow* 153.4ms +/- 0.4% 154.6ms +/- 0.7% significant tagcloud: *1.04x as slow* 113.6ms +/- 0.6% 118.0ms +/- 0.7% significant unpack-code: *1.01x as slow* 114.4ms +/- 1.2% 116.0ms +/- 0.8% significant validate-input: *1.04x as slow* 101.2ms +/- 0.5% 105.6ms +/- 2.4% significant

Alexey Proskuryakov

Comment 2 2008-05-27 09:46:13 PDT

Progressing memory fragmentation is what comes to mind first, but ideally, this shouldn't happen, of course.

Alexey Proskuryakov

Comment 3 2008-06-03 05:27:22 PDT