Latency numbers every programmer should know

latency.markdown

Latency numbers every programmer should know

L1 cache reference ......................... 0.5 ns
Branch mispredict ............................ 5 ns
L2 cache reference ........................... 7 ns
Mutex lock/unlock ........................... 25 ns
Main memory reference ...................... 100 ns             
Compress 1K bytes with Zippy ............. 3,000 ns  =   3 µs
Send 2K bytes over 1 Gbps network ....... 20,000 ns  =  20 µs
SSD random read ........................ 150,000 ns  = 150 µs
Read 1 MB sequentially from memory ..... 250,000 ns  = 250 µs
Round trip within same datacenter ...... 500,000 ns  = 0.5 ms
Read 1 MB sequentially from SSD* ..... 1,000,000 ns  =   1 ms
Disk seek ........................... 10,000,000 ns  =  10 ms
Read 1 MB sequentially from disk .... 20,000,000 ns  =  20 ms
Send packet CA->Netherlands->CA .... 150,000,000 ns  = 150 ms

Assuming ~1GB/sec SSD

Visual chart provided by ayshen

Data by Jeff Dean

Originally by Peter Norvig

latency_humanized.markdown

Lets multiply all these durations by a billion:

Magnitudes:

Minute:

L1 cache reference                  0.5 s         One heart beat (0.5 s)
Branch mispredict                   5 s           Yawn
L2 cache reference                  7 s           Long yawn
Mutex lock/unlock                   25 s          Making a coffee

Hour:

Main memory reference               100 s         Brushing your teeth
Compress 1K bytes with Zippy        50 min        One episode of a TV show (including ad breaks)

Day:

Send 2K bytes over 1 Gbps network   5.5 hr        From lunch to end of work day

Week

SSD random read                     1.7 days      A normal weekend
Read 1 MB sequentially from memory  2.9 days      A long weekend
Round trip within same datacenter   5.8 days      A medium vacation
Read 1 MB sequentially from SSD    11.6 days      Waiting for almost 2 weeks for a delivery

Year

Disk seek                           16.5 weeks    A semester in university
Read 1 MB sequentially from disk    7.8 months    Almost producing a new human being
The above 2 together                1 year

Decade

Send packet CA->Netherlands->CA     4.8 years     Average time it takes to complete a bachelor's degree

25s is making one coffee. 7.8 months is a tad faster than producing a new human being. One disk-seek (16 weeks) and then reading 1MiB (7.8 months) is one year.

25s is making one coffee. 7.8 months is a tad faster than producing a new human being. One disk-seek (16 weeks) and then reading 1MiB (7.8 months) is one year.

thx, will incorporate those!

thx, will incorporate those!

16.5 weeks ~= four months ~= full semester at university ~= summer break ?

16.5 weeks ~= four months ~= full semester at university ~= summer break ?

branch mis-predict: a yawn.

Such an accurate portrayal. "yawwwn aw damn it fell through to the else again."

branch mis-predict: a yawn.

Such an accurate portrayal. "yawwwn aw damn it fell through to the else again."

4.8 years is very roughly two round trips to Mars. Or a one-way trip to Europa.

4.8 years is very roughly two round trips to Mars. Or a one-way trip to Europa.

neat, thanks!

neat, thanks!

4.8 years is about how long a bachelor's degree takes on average (4.7 years in 2008 according to this page).

4.8 years is about how long a bachelor's degree takes on average (4.7 years in 2008 according to this page).

Should add date of creation to text itself (as of ...)

Should add date of creation to text itself (as of ...)

If enough people start reading this, the last one will have to wait several years

If enough people start reading this, the last one will have to wait several years

Here's a tool to visualize these numbers over time: http://www.eecs.berkeley.edu/~rcs/research/interactive_latency.html

Here's a tool to visualize these numbers over time: http://www.eecs.berkeley.edu/~rcs/research/interactive_latency.html

Great interactive chart! Suggest to add CPU cycles for a simple single instructions such as ADD, MOV, MULT (e.g. "multiplying two 16bit integers".

Great interactive chart! Suggest to add CPU cycles for a simple single instructions such as ADD, MOV, MULT (e.g. "multiplying two 16bit integers".

I'd like to see a duration for function-call overhead. Late 80s when I was in school, it was several micro-seconds, enough that people had concerns about having unnecessary function calls. I'm sure that doesn't apply any more, but I don't have real numbers.

I'd like to see a duration for function-call overhead. Late 80s when I was in school, it was several micro-seconds, enough that people had concerns about having unnecessary function calls. I'm sure that doesn't apply any more, but I don't have real numbers.

@legrady it depends on a lot of factors. Is it a virtual function call?

@legrady it depends on a lot of factors. Is it a virtual function call?

大神

大神

If reading 1MB from an SSD costs 1ms, what would the cost be to read 10MB sequentially from an SSD?

If reading 1MB from an SSD costs 1ms, what would the cost be to read 10MB sequentially from an SSD?

This is a great collection. I just dont get where or how i am able to make coffee in just 25s.

This is a great collection. I just dont get where or how i am able to make coffee in just 25s.

Agree @b1nary . if someone knows how to do that, please share the source code 😄

Agree @b1nary . if someone knows how to do that, please share the source code 😄

That would be a shocker if devops status page turned into humanized numbers one day (sometime in april).
We should all start working harder to improve our numbers ! and enjoy more Round trip within same datacenter 😃

That would be a shocker if devops status page turned into humanized numbers one day (sometime in april).
We should all start working harder to improve our numbers ! and enjoy more Round trip within same datacenter 😃

Do not forget:

3ms: Time till a wrongly configured sendmail timeouts and fails to deliver a mail. Roughly corresponds to mail servers in a 500km (3 millilightseconds) radius

6h: Time to send a mail across those 500km via RFC 1149

Do not forget:

3ms: Time till a wrongly configured sendmail timeouts and fails to deliver a mail. Roughly corresponds to mail servers in a 500km (3 millilightseconds) radius

6h: Time to send a mail across those 500km via RFC 1149

cool

cool

@stultus @b1nary we have a coffee machine that makes coffee. Ta-Dah! 😄

@stultus @b1nary we have a coffee machine that makes coffee. Ta-Dah! 😄

good, it's cool!

good, it's cool!

cool! great summary

cool! great summary

this is awesome. Thank you

this is awesome. Thank you

hmm:

branch misprediction penalty on Haswell ~ 1500 ns vs 5 ns in the gist. That's three orders of magnitude of error

EDIT:

I used the ticks from Windows (are 10K in a ms) which is incorrect related to the gist.

If the Haswell CPU is running 3.6Ghz - one cycle would equal to 0.27ns that would mean a branch miss would be 4.05ns - seems about right now.

hmm:

branch misprediction penalty on Haswell ~ 1500 ns vs 5 ns in the gist. That's three orders of magnitude of error

EDIT:

I used the ticks from Windows (are 10K in a ms) which is incorrect related to the gist.

If the Haswell CPU is running 3.6Ghz - one cycle would equal to 0.27ns that would mean a branch miss would be 4.05ns - seems about right now.

cool..... thank you

cool..... thank you

Multiplying by a billion stretches the timescales out too much for my taste. So I came up with an alternate list based on multiplying by only 22,000:

L1 cache reference ..................  0.000011 seconds (SR-71 travels 1cm)
Branch mispredict ...................  0.000110 sec (Bullet travels 4cm)
L2 cache reference ..................  0.000154 sec (Boeing 777 travels 4cm)
Mutex lock/unlock ...................  0.00055 sec (Time before you hear a fingersnap made in front of your face [speed of sound across 19cm])
Main memory reference ...............  0.0022 sec (Camera shutter on a sunny day [1/400 - 1/500 shutter speed])
Compress 1K bytes with Zippy ........  0.066 sec (Lightning bolt travels 4km from cloud to ground)
Send 2K bytes over 1 Gbps network ...  0.44 sec (Fastball from pitcher to home plate)
SSD random read .....................  3.3 sec (SR-71 travels 3.1km)
Read 1 MB sequentially from memory ..  5.5 sec (Yawn)
Round trip within same datacenter ... 11.0 sec (A Cheetah runs 200m)
Read 1 MB sequentially from SSD* .... 22.0 sec (Usain Bolt runs 200m)
Disk seek ...........................  3.6 minutes (Brewing coffee in a French Press)
Read 1 MB sequentially from disk ....  7.3 min (A performance of the first movement of Beethoven's 5th Symphony)
Send packet CA->Netherlands->CA ..... 55.0 min (Going for a brisk 5km walk)

Multiplying by a billion stretches the timescales out too much for my taste. So I came up with an alternate list based on multiplying by only 22,000:

L1 cache reference ..................  0.000011 seconds (SR-71 travels 1cm)
Branch mispredict ...................  0.000110 sec (Bullet travels 4cm)
L2 cache reference ..................  0.000154 sec (Boeing 777 travels 4cm)
Mutex lock/unlock ...................  0.00055 sec (Time before you hear a fingersnap made in front of your face [speed of sound across 19cm])
Main memory reference ...............  0.0022 sec (Camera shutter on a sunny day [1/400 - 1/500 shutter speed])
Compress 1K bytes with Zippy ........  0.066 sec (Lightning bolt travels 4km from cloud to ground)
Send 2K bytes over 1 Gbps network ...  0.44 sec (Fastball from pitcher to home plate)
SSD random read .....................  3.3 sec (SR-71 travels 3.1km)
Read 1 MB sequentially from memory ..  5.5 sec (Yawn)
Round trip within same datacenter ... 11.0 sec (A Cheetah runs 200m)
Read 1 MB sequentially from SSD* .... 22.0 sec (Usain Bolt runs 200m)
Disk seek ...........................  3.6 minutes (Brewing coffee in a French Press)
Read 1 MB sequentially from disk ....  7.3 min (A performance of the first movement of Beethoven's 5th Symphony)
Send packet CA->Netherlands->CA ..... 55.0 min (Going for a brisk 5km walk)

@Kevin-Hamilton There is a reason to stretch it out that much. From a human perspective, it is really hard to do anything in less than a second. As such, the ridiculously long times give you a better idea of what a computer might be able to do during that disk seek, if it weren't waiting for that disk seek.

Because of this, most of the similar pages I have seen use something like 1 second for a clock cycle (so multiply everything there by 3 or 4). It gives a really good sense of machine sympathy.

@Kevin-Hamilton There is a reason to stretch it out that much. From a human perspective, it is really hard to do anything in less than a second. As such, the ridiculously long times give you a better idea of what a computer might be able to do during that disk seek, if it weren't waiting for that disk seek.

Because of this, most of the similar pages I have seen use something like 1 second for a clock cycle (so multiply everything there by 3 or 4). It gives a really good sense of machine sympathy.

Last year, I came up with this concept for an infographic illustrating these latency numbers with time analogies (if 1 CPU cycle = 1 second). Here was the result (attached, and here's a link: http://imgur.com/8LIwV4C)

Last year, I came up with this concept for an infographic illustrating these latency numbers with time analogies (if 1 CPU cycle = 1 second). Here was the result (attached, and here's a link: http://imgur.com/8LIwV4C)

Excellent idea ! Great page !
Perhaps an interesting comparison:

• the human eye requires 13 ms to identify an image
• an eye blink is around 100 ms
• the reaction time from eye to mouse is around 215 ms

Excellent idea ! Great page !
Perhaps an interesting comparison:

• the human eye requires 13 ms to identify an image
• an eye blink is around 100 ms
• the reaction time from eye to mouse is around 215 ms

Cool. Loved it.

Cool. Loved it.

very interesting!

very interesting!

Excelent Gist.

Excelent Gist.