Cloud computing systems today, whether open-source or used inside companies, are built using a common set of core techniques, algorithms, and design philosophies – all centered around distributed systems. Learn about such fundamental distributed computing "concepts" for cloud computing. Some of these concepts include: clouds, MapReduce, key-value/NoSQL stores, classical distributed algorithms, widely-used distributed algorithms, scalability, trending areas, and much, much more! Know how these systems work from the inside out. Get your hands dirty using these concepts with provided homework exercises. In the programming assignments, implement some of these concepts in template code (programs) provided in the C++ programming language. Prior experience with C++ is required. The course also features interviews with leading researchers and managers, from both industry and academia.
2.3. Gossip-Style Membership
Loading...
From the course by University of Illinois at Urbana-Champaign
Cloud Computing Concepts, Part 1
579 ratings
University of Illinois at Urbana-Champaign
579 ratings
Course 1 of 6 in the Specialization Cloud Computing
From the lesson
Week 2: Gossip, Membership, and Grids
Lesson 1: This module teaches how the multicast problem is solved by using epidemic/gossip protocols. It also teaches analysis of such protocols. Lesson 2: This module covers the design of failure detectors, a key component in any distributed system. Membership protocols, which use failure detectors as components, are also covered. Lesson 3: This module covers Grid computing, an important precursor to cloud computing.
Meet the Instructors
Indranil GuptaProfessor
Department of Computer Science
In this lecture we will see how gossip-style, uh,
failure detection works.
Gossip-style failure detection is essentially, ah,
a variant of call-to-all heartbeating
which is more robust than all-to-all heartbeating.
Once again, the basic concept is that
every process, uh, wants to send out it-its heartbeats
to all the other processes in the system.
The heartbeats are again incremented sequence numbers
that are incremented locally.
Processes timeout waiting for heartbeats
and mark the corresponding sources as having failure.
We want really good accuracy properties here,
which was lacking in the all to all variant of heartbeating.
So here is how gossip-style failure detection works
in a nutshell.
Here I have an example with, uh, four processes, uh,
marked as 1, 2, 3, and 4.
Every process maintains a table.
Let's look at the table for 1,
and the table for 1 shows, uh, four entries
for each of the four processes in the system.
There are, uh, so, four rows.
There are three columns.
The first column is the address and the process
which is basically the ID.
The second is the heartbeat counter received
from that corresponding process.
Third is the local time at which
that heartbeat counter was last updated.
So, for instance, at process 1,
the row for 2 says that the last heartbeat that was received
from process 2 at process 1 was numbered 10, 10, 3,
and it was received when the local time at process 1 was 62.
The local time is needed essentially
because times may be unsynchronized
across different processes
as we have discussed before,
and this is an asynchronous system after all,
and so we wanna be using the local time
to mark when the heartbeat was last updated.
Periodically, each process sends over this entire table
to a few of its neighbors selected at random.
Okay, this is known as, uh, gossip.
So nodes are processes
periodically gossip their membership list.
So, for instance, if 1 selects 2 at random
and sends it its membership list,
say 2's membership looks like this, right.
What 2 does when it receives 1's membership list
is that it merges it into its own list.
It merges it row by row;
so for instance, it looks at row number 1
and it says that hey, my, uh, latest for heartbeat
for, uh, 1 is 10,118,
but I am receiving a heartbeat for 1 which is 10,120,
which is later,
and so I am going to update that, uh, heartbeat
to be 10,120.
Also I am going to update the, uh, time
component of that particular row
from 64 to my current local time
which happens to be 70 at this node.
For the second row, it doesn't update it
because it has a low- uh, a later heartbeat.
After all it was in row number 2.
Uh, for the third row, for node number 3,
again it does update it,
uh, because, uh, 1's incoming heartbeat, uh,
table has a-a higher heartbeat counter,
so it becomes 10,098 inherited from 1,
and the local time is marked at 70 over here.
Number 4 is not updated because the heartbeat counters
are the same here, uh, both at 1 and 2, so that's not updated.
It is kept as is.
Essentially this, ah, is the way, ah, gossip works.
Ah, the essential component here is that, ah, 1, ah,
periodically selects, um, uh, a few other members at, uh, random
from its membership list
and sends it copies of, uh, its membership list
and they do the protocol here.
The timeout works the same way.
If, ah, there is a threshold typically for, ah, timing out
and, um, when a-a particular row's, um, entry
was last updated more than that timeout, uh, seconds ago
you mark that node as our processes having failed.
So the heartbeat has not increased
for more than T failed seconds.
Notice that this is a local time
at the, uh, process that is maintaining the table,
then the member is marked as having failed.
However, you don't delete that particular entry right away
from the row, uh, from the table that you have.
Uh, you wait for another T cleanup seconds.
Typically cleanup is also around the same as a T fail,
and only after that do you delete the member
from your list.
So why do we have two different timeouts?
Why not delete the member right away?
Well you could have, uh, this very interesting phenomenon
if you delete the member right away
where an entry might never go away.
Consider here this particular scenario
where process 1 has an entry for 3 marked as, uh,
a heartbeat 10,098 with local time 55
and process 2 also has a similar heart- uh, entry for process 3.
Process 1 marks- uh, sorry, process 2 marks row 3
as having failed because the current timeout is 75
and the T fail is 24 seconds ago,
so the timeout just elapsed;
however, process 1, uh, has not yet reached that timeout of 24
and so it has not deleted 3 yet,
but 2 has gone ahead and deleted 3.
Soon enough 1 sends its heartbeat list over to 2
and 2 sees the entry 3- for 3 in there and says,
"Oh, this is a new, ah, membership list entry.
"I'm gonna add it back in.
"And in fact, I'm gonna add it back in
"with the new local time, uh, at node 2."
And so essentially you can see what's going on here,
that after a while 1 will delete the entry for 3,
but 2 might select, uh, 1 as the gossip target
and it might send it 3's entry and then 1 will add it back,
and so because of this ping pong behavior that happens,
3's entry might never, ever go away from the system.
So it is to avoid exactly this
that you remember, uh, the entry for another T cleanup seconds
typically the same as T fail timeout
and if you receive another entry for the same, uh, process,
you don't pay heed to it.
You ignore it.
So, let's go back to the discussion of, uh,
what happens if the gossip period, uh,
T gossip, is decreased.
Um, so once again, uh, the gossip property says that
a single heartbeat, uh, takes O(log(N)) time to propagate
as long as you have enough bandwidth, uh,
to, uh, send out the entire, uh, membership list,
uh, every time, uh, you gossip.
Uh, so, uh, N heartbeats take O(log(N)) time to propagate
and the bandwidth allowed per node, ah,
is allowed to be O(N).
Uh, however, the bandwidth allowed is O(1),
you're allowed to send out only a few sampled rows
from your membership list,
theyn the-then the, um, uh, then the, um, dissemination time
might go up to Order(N-log(N)).
Essentially this is, uh, inversely proportionate.
In the gossip period, T gossip is decreased;
essentially what happens, ah,
is that you have a higher bandwidth, uh,
and that, uh, you spre-you send out gossips, uh, much quicker
so you can have a shorter timeout
for the failure detection itself.
So essentially, the T gossip is a tradeoff
between bandwidth and the detection time.
Uh, the bandwidth again, uh, um, determines, uh, um,
the detection timeouts.
The lower your bandwidth limits are, uh,
the longer y-y-your detection time has to be.
So I encourage you to think about what happens
if you have O(K) bandwidth in, uh, this.
Once again, if you increase T_fail and T_cleanup,
the timeouts there,
what happens to the false positive, uh, probability?
Well essentially what happens
is that your detection time increases,
but then at the same time, uh,
you have a-a lower false positive rate
because you give, uh, non-faulty nodes
that are mistakenly detected slightly longer
for the heartbeats to make it across.
So essentially you have a bunch of knobs here.
Uh, the T_gossip is one of those knobs and the timeouts,
the T_fail and the T_cleanup are the other knobs,
and the bandwidth constraints, uh, um, are the other knobs
which help you to trade off between the false positive rate,
uh, the detection time, and the bandwidth
that is, uh, used by your gossip protocol itself.
So in the next lecture we'll see whether or not
this in is in fact optimal.
Is there an optimal failure detector?
How close d-do the failure detectors
we have discussed so far come to this optimal, uh, amount?
Explore our Catalog
Join for free and get personalized recommendations, updates and offers.
Coursera provides universal access to the world’s best education,
partnering with top universities and organizations to offer courses online.
© 2018 Coursera Inc. All rights reserved.
