On Saturday, November 28, 2015, Yeongjin Jang yeongjinjanggrad@gmail.com wrote:

Thank you for your comments!

On Sat, Nov 28, 2015 at 2:33 PM, Brian Wolff bawolff@gmail.com wrote:

...

On 11/28/15, Yeongjin Jang yeongjinjanggrad@gmail.com wrote:

...

Hi,

I am Yeongjin Jang, a Ph.D. Student at Georgia Tech.

In our lab (SSLab, https://sslab.gtisc.gatech.edu/), we are working on a project called B2BWiki, which enables users to share the contents of Wikipedia through WebRTC (peer-to-peer sharing).

Website is at here: http://b2bwiki.cc.gatech.edu/

The project aims to help Wikipedia by donating computing resources from the community; users can donate their traffic (by P2P

communication)

...

...

and storage (indexedDB) to reduce the load of Wikipedia servers. For larger organizations, e.g. schools or companies that have many local users, they can donate a mirror server similar to GNU FTP servers, which can bootstrap peer sharing.

Potential benefits that we think of are following.

1. Users can easily donate their resources to the community.

Just visit the website.

2. Users can get performance benefit if a page is loaded from

multiple local peers / local mirror (page load time got faster!).

3. Wikipedia can reduce its server workload, network traffic, etc.

4. Local network operators can reduce network traffic transit

(e.g. cost that is caused by delivering the traffic to the outside).

While we are working on enhancing the implementation, we would like to ask the opinions from actual developers of Wikipedia. For example, we want to know whether our direction is correct or not (will it actually reduce the load?), or if there are some other

concerns

...

...

that we missed, that can potentially prevent this system from working as intended. We really want to do somewhat meaningful work that actually helps run Wikipedia!

Please feel free to give as any suggestions, comments, etc. If you want to express your opinion privately, please contact sslab@cc.gatech.edu.

Thanks,

--- Appendix ---

I added some detailed information about B2BWiki in the following.

# Accessing data When accessing a page on B2BWiki, the browser will query peers first.

1. If there exist peers that hold the contents, peer to peer download

happens. 2) otherwise, if there is no peer, client will download the content from the mirror server. 3) If mirror server does not have the content, it downloads from Wikipedia server (1 access per first download, and update).

# Peer lookup To enable content lookup for peers, we manage a lookup server that holds a page_name-to-peer map. A client (a user's browser) can query the list of peers that currently hold the content, and select the peer by its freshness (has hash/timestamp of the content, has top 2 octet of IP address (figuring out whether it is local peer or not), etc.

# Update, and integrity check Mirror server updates its content per each day (can be configured to update per each hour, etc). Update check is done by using If-Modified-Since header from Wikipedia server. On retrieving the content from Wikipedia, the mirror server stamps a timestamp and sha1 checksum, to ensure the freshness of data and its integrity. When clients lookup and download the content from the peers, client will compare the sha1 checksum of data with the checksum from lookup server.

In this settings, users can get older data (they can configure how to tolerate the freshness of data, e.g. 1day older, 3day, 1 week older, etc.), and the integrity is guaranteed by mirror/lookup server.

More detailed information can be obtained from the following website.

http://goo.gl/pSNrjR (URL redirects to SSLab@gatech website)

Please feel free to give as any suggestions, comments, etc.

Thanks,

Yeongjin Jang

---

Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l

Hi,

This is some interesting stuff, and I think research along these lines (That is, leveraging webrtc to deliver content in a P2P manner over web browsers) will really change the face of the internet in the years to come.

As for Wikipedia specifically (This is all just my personal opinion. Others may disagree with me):

*Wikipedia is a a fairly stable/mature site. I think we're past the point where its a good idea to experiment with experimental technologies (Although mirrors of Wikipedia are a good place to do that). We need stability and proven technology.

That's true. Our current prototype works well for the testing, but not sure for the robustness in the wild, yet. We want to develop this to be more stable.

...

*Bandwidth makes up a very small portion of WMF's expenses (I'm not sure how small. Someone once told me that donation processing costs takes up more money then raw bandwidth costs. Don't know if that's true, but bandwidth is certainly not the biggest expense).

Your scheme primarily serves to offload bandwidth of cached content to other people. But serving cached content (by which I mean, anonymous users getting results from varnish) is probably the cheapest (in terms of computational resources) part of our setup. The hard part is things like parsing wikitext->html, and otherwise generating pages.

Yes, it can be. But for the network traffic bandwidth, we think that it could benefit both side (an organizations that runs B2BWiki and Wikipedia), because it would reduce not only the traffic that hits Wikipedia, but also egress traffic from the LAN (or ISP) to Wikipedia.

And we know that this is just a hypothesis, so we want to do analysis on potential reduction in traffic/cost with network data stats.

Is there any point that I can get stats for the bandwidth, such as daily traffic for serving Wikipedia servers, etc? (Please let me know if you know any point)

Maybe the (third column of) projectcounts files at http://dumps.wikimedia.org/other/pagecounts-raw/2015/2015-11/

Financial cost of bandwidth would probably be in one of the annual plan documents somewhere.

I visited several stats pages, such as

https://stats.wikimedia.org/EN/ChartsWikipediaEN.htm https://en.wikipedia.org/wiki/Wikipedia:Statistics

, and those sites let me know about how many page accesses, edits were happened but not for the traffic, while ganglia site gave me too finer grained stats.

And what we got from our local network data is, from Georgia Tech, it shows that 50GB of download per day from Wikipedia (including wikipedia.org and upload.wikimedia.org, based on source IP address). From the simple calculations, the cost of 18TB / year will be around $600, for serving 30K person organizations.

...

*24 hours to page update is generally considered much too slow (Tolerance for anons is probably a bit higher than logged in users, but still). People expect their changes to appear for everyone, immediately. We want the delay to be in seconds, not days. I think its unlikely that any sort of expire scheme would be acceptable. We need active cache invalidation upon edit.

Yes, we have thought about cache expiring/update/invalidation, for example, whenever a peer detects update, propagate the update to its neighbor to update/invalidate the cache. We did not concentrate on that part much, as we had similar thoughts that non-editors can be more tolerable. The scheme is not in there yet, but it is worth to implement if editors care it much.

...

*Lack of analysis of scalability (I just briefly skimmed the google cache version of the page you linked [your webserver had the connection keep being reset], so its possible I missed this). I didn't see any analysis of how your system scales with load. Perhaps that's because you're still in the development stage, and the design isn't finalized(?) Anyways, scalability analysis is important when we're talking about wikipedia. Does this design still work if you have 100,000 (or even more) peers?

We think scalability is very important for B2BWiki. While we were doing micro-benchmark on the servers of B2BWiki, we realized that current one cannot support more than 10K concurrent peers. We've been update the internal structure from September, and now we are targeting support up to 100K peers (we think this would be enough for supporting a metropolitan local area).

Implementation with new data structure will be available very soon, I hope we can show the good result that shows its scalability.

The other scalability concern would be for obscure articles. I havent really looked at your code, so maybe you cover it - but wikipedia has over 5 million articles (and a lot more when you count non-content pages). The group of peers is presumably going to have high churn (since they go away when you browse somewhere else). Id worry the overhead of keeping track of which peer knows what, especially given how fast the peers change to be a lot. I also expect that for lots of articles, only a very small number of peers will know them.

...

*Privacy concerns - Would a malicious person be able to force themselves to be someone's preferred peer, and spy on everything they read, etc.

*DOS concerns - Would a malicious peer or peers be able to prevent an honest user from connecting to the website? (I didn't look in detail at how you select peers and handle peer failure, so I'm not sure if this applies)

Nice points! For privacy, we want to implement k-anonymity scheme on the page access. However, it incurs more bandwidth consumption and potential performance overhead on the system.

Malicious peers can act as if they hold legitimate content (while actually not), or making null request to the peers. We are currently thinking about black-listing such malicious peers, and live-migration of mirror/peer servers if they fails, but more fundamental remedy is required.

--

...

Anyways, I think this sort of think is interesting, but I think your system would be more suited to people running a small static website, that want to scale to very high numbers, rather than Wikipedia's use case.

-- -bawolff

---

Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l

-- Yeongjin Jang _______________________________________________ Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l

On Sun, Nov 29, 2015 at 1:59 PM, Brian Wolff bawolff@gmail.com wrote:

On Saturday, November 28, 2015, Yeongjin Jang yeongjinjanggrad@gmail.com wrote:

...

Thank you for your comments!

On Sat, Nov 28, 2015 at 2:33 PM, Brian Wolff bawolff@gmail.com wrote:

...

On 11/28/15, Yeongjin Jang yeongjinjanggrad@gmail.com wrote:

...

Hi,

I am Yeongjin Jang, a Ph.D. Student at Georgia Tech.

In our lab (SSLab, https://sslab.gtisc.gatech.edu/), we are working on a project called B2BWiki, which enables users to share the contents of Wikipedia through WebRTC (peer-to-peer sharing).

Website is at here: http://b2bwiki.cc.gatech.edu/

The project aims to help Wikipedia by donating computing resources from the community; users can donate their traffic (by P2P

communication)

...

...

...

and storage (indexedDB) to reduce the load of Wikipedia servers. For larger organizations, e.g. schools or companies that have many local users, they can donate a mirror server similar to GNU FTP servers, which can bootstrap peer sharing.

Potential benefits that we think of are following.

1. Users can easily donate their resources to the community.

Just visit the website.

2. Users can get performance benefit if a page is loaded from

multiple local peers / local mirror (page load time got faster!).

3. Wikipedia can reduce its server workload, network traffic, etc.

4. Local network operators can reduce network traffic transit

(e.g. cost that is caused by delivering the traffic to the outside).

While we are working on enhancing the implementation, we would like to ask the opinions from actual developers of Wikipedia. For example, we want to know whether our direction is correct or not (will it actually reduce the load?), or if there are some other

concerns

...

...

...

that we missed, that can potentially prevent this system from working as intended. We really want to do somewhat meaningful work that actually helps run Wikipedia!

Please feel free to give as any suggestions, comments, etc. If you want to express your opinion privately, please contact sslab@cc.gatech.edu.

Thanks,

--- Appendix ---

I added some detailed information about B2BWiki in the following.

# Accessing data When accessing a page on B2BWiki, the browser will query peers first.

1. If there exist peers that hold the contents, peer to peer download

happens. 2) otherwise, if there is no peer, client will download the content from the mirror server. 3) If mirror server does not have the content, it downloads from Wikipedia server (1 access per first download, and update).

# Peer lookup To enable content lookup for peers, we manage a lookup server that holds a page_name-to-peer map. A client (a user's browser) can query the list of peers that currently hold the content, and select the peer by its freshness (has hash/timestamp of the content, has top 2 octet of IP address (figuring out whether it is local peer or not), etc.

# Update, and integrity check Mirror server updates its content per each day (can be configured to update per each hour, etc). Update check is done by using If-Modified-Since header from Wikipedia server. On retrieving the content from Wikipedia, the mirror server stamps a timestamp and sha1 checksum, to ensure the freshness of data and its integrity. When clients lookup and download the content from the peers, client will compare the sha1 checksum of data with the checksum from lookup server.

In this settings, users can get older data (they can configure how to tolerate the freshness of data, e.g. 1day older, 3day, 1 week older, etc.), and the integrity is guaranteed by mirror/lookup server.

More detailed information can be obtained from the following website.

http://goo.gl/pSNrjR (URL redirects to SSLab@gatech website)

Please feel free to give as any suggestions, comments, etc.

Thanks,

Yeongjin Jang

---

Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l

Hi,

This is some interesting stuff, and I think research along these lines (That is, leveraging webrtc to deliver content in a P2P manner over web browsers) will really change the face of the internet in the years to come.

As for Wikipedia specifically (This is all just my personal opinion. Others may disagree with me):

*Wikipedia is a a fairly stable/mature site. I think we're past the point where its a good idea to experiment with experimental technologies (Although mirrors of Wikipedia are a good place to do that). We need stability and proven technology.

That's true. Our current prototype works well for the testing, but not sure for the robustness in the wild, yet. We want to develop this to be more stable.

...

*Bandwidth makes up a very small portion of WMF's expenses (I'm not sure how small. Someone once told me that donation processing costs takes up more money then raw bandwidth costs. Don't know if that's true, but bandwidth is certainly not the biggest expense).

Your scheme primarily serves to offload bandwidth of cached content to other people. But serving cached content (by which I mean, anonymous users getting results from varnish) is probably the cheapest (in terms of computational resources) part of our setup. The hard part is things like parsing wikitext->html, and otherwise generating pages.

Yes, it can be. But for the network traffic bandwidth, we think that it could benefit

both

...

side (an organizations that runs B2BWiki and Wikipedia), because it would reduce not only the traffic that hits Wikipedia, but also egress traffic from the LAN (or ISP) to Wikipedia.

And we know that this is just a hypothesis, so we want to do analysis on potential reduction in traffic/cost with network data stats.

Is there any point that I can get stats for the bandwidth, such as daily traffic for serving Wikipedia servers, etc? (Please let me know if you know any point)

Maybe the (third column of) projectcounts files at http://dumps.wikimedia.org/other/pagecounts-raw/2015/2015-11/

Financial cost of bandwidth would probably be in one of the annual plan documents somewhere.

Thank you for the pointer!

I presume that only counts HTML data, since the counted number is far smaller than mediacount stat result.

From projectcounts, traffic on the day of 11/01/2015 is like

around 4.4TB for en.wikipedia (including mobile), and around 7TB for overall languages.

And, from the mediacount stats at http://dumps.wikimedia.org/other/mediacounts/ There was 110TB downloads referred within WMF sites on that day.

I recall that I saw financial statement of WMF that states around $2.3M was spent for Internet Hosting. I am not sure whether it includes management cost for computing resources (server clusters such as eqiad) or not.

Not sure following simple calculation works; 117 TB per day, for 365 days, if $0.05 per GB, then it is around $2.2M. Maybe it would be more accurate if I contact analytics team directly.

...

I visited several stats pages, such as

https://stats.wikimedia.org/EN/ChartsWikipediaEN.htm https://en.wikipedia.org/wiki/Wikipedia:Statistics

, and those sites let me know about how many page accesses, edits were happened but not for the traffic, while ganglia site gave me too finer grained stats.

And what we got from our local network data is, from Georgia Tech, it shows that 50GB of download per day from Wikipedia (including wikipedia.org and upload.wikimedia.org, based on source IP address). From the simple calculations, the cost of 18TB / year will be around $600, for serving 30K person organizations.

...

*24 hours to page update is generally considered much too slow (Tolerance for anons is probably a bit higher than logged in users, but still). People expect their changes to appear for everyone, immediately. We want the delay to be in seconds, not days. I think its unlikely that any sort of expire scheme would be acceptable. We need active cache invalidation upon edit.

Yes, we have thought about cache expiring/update/invalidation, for example, whenever a peer detects update, propagate the update to its neighbor to update/invalidate the cache. We did not concentrate on that part much, as we had similar thoughts that non-editors can be more tolerable. The scheme is not in there yet, but it is worth to implement if editors care it much.

...

*Lack of analysis of scalability (I just briefly skimmed the google cache version of the page you linked [your webserver had the connection keep being reset], so its possible I missed this). I didn't see any analysis of how your system scales with load. Perhaps that's because you're still in the development stage, and the design isn't finalized(?) Anyways, scalability analysis is important when we're talking about wikipedia. Does this design still work if you have 100,000 (or even more) peers?

We think scalability is very important for B2BWiki. While we were doing micro-benchmark on the servers of B2BWiki, we realized that current one cannot support more than 10K concurrent peers. We've been update the internal structure from September, and now we are targeting support up to 100K peers (we think this would be enough for supporting a metropolitan local area).

Implementation with new data structure will be available very soon, I hope we can show the good result that shows its scalability.

The other scalability concern would be for obscure articles. I havent really looked at your code, so maybe you cover it - but wikipedia has over 5 million articles (and a lot more when you count non-content pages). The group of peers is presumably going to have high churn (since they go away when you browse somewhere else). Id worry the overhead of keeping track of which peer knows what, especially given how fast the peers change to be a lot. I also expect that for lots of articles, only a very small number of peers will know them.

That's true. Dynamically registering / un-registering lookup table gives high overhead on the servers (in both computation & memory usage). Distributed solutions like DHT is there, but we think there could be a trade-off on lookup time for using de-centralized (managing lookup table on the server) versus fully distributed architecture (DHT).

Our prior "naive" implementation costs like if each user has 5K pages cached (with around 50K images), when 10K concurrent user presents it consumes around 35GB of memory, and each registering incurs 500K bytes of network traffic.

We thought it is not that useful, so now we are trying to come up with more lightweight implementation. We hope to have a practically meaningful micro-benchmark result on the new implementation.

...

...

*Privacy concerns - Would a malicious person be able to force themselves to be someone's preferred peer, and spy on everything they read, etc.

*DOS concerns - Would a malicious peer or peers be able to prevent an honest user from connecting to the website? (I didn't look in detail at how you select peers and handle peer failure, so I'm not sure if this applies)

Nice points! For privacy, we want to implement k-anonymity scheme on the page access. However, it incurs more bandwidth consumption and potential performance overhead on the system.

Malicious peers can act as if they hold legitimate content (while actually not), or making null request to the peers. We are currently thinking about black-listing such malicious peers, and live-migration of mirror/peer servers if they fails, but more fundamental remedy is required.

--

...

Anyways, I think this sort of think is interesting, but I think your system would be more suited to people running a small static website, that want to scale to very high numbers, rather than Wikipedia's use case.

-- -bawolff

---

Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l

-- Yeongjin Jang _______________________________________________ Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l

---

Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l

On 30.11.2015 20:47, Ryan Lane wrote:

On Sun, Nov 29, 2015 at 8:18 PM, Yeongjin Jang yeongjinjanggrad@gmail.com wrote:

...

I recall that I saw financial statement of WMF that states around $2.3M was spent for Internet Hosting. I am not sure whether it includes management cost for computing resources (server clusters such as eqiad) or not.

That's the cost for datacenters, hardware, bandwidth, etc..

...

Not sure following simple calculation works; 117 TB per day, for 365 days, if $0.05 per GB, then it is around $2.2M. Maybe it would be more accurate if I contact analytics team directly.

That calculation doesn't work because it doesn't take into account peering agreements, or donated (or heavily discounted) transit contracts. Bandwidth is one of the cheaper overall costs.

Something your design doesn't take into account for bandwidth costs is that the world is trending to mobile and mobile bandwidth costs are generally very high. It's likely this p2p approach will be many orders of magnitude more expensive than the current approach.

A decentralized approach doesn't benefit from the economics of scale. Instead of being able to negotiate transit pricing and eliminating cost through peering, you're externalizing the cost at the consumer rate, which is the highest possible rate.

While that is often true, there are notable exception, growing both in scale and number.

a) We have campus situations where a large university, company, or public agency with tens or hundreds of thousands of peers run a network that they pay for anyways, that is needed for peers to connect to Wiki* anyways, and that is available to peers at no (additional) cost. While external traffic cost are of relatively little concern, quick response times often are, especially in classroom situations where up to several hundred students may look at the same articles virtually at once. b) There is a fast growing international movement for open wireless radio networks (freifunk) of comparatively low bandwidth in neighborhoods. Those can benefit a lot from local peering, imho.

Purodha

...

The other scalability concern would be for obscure articles. I havent really looked at your code, so maybe you cover it - but wikipedia has over 5 million articles (and a lot more when you count non-content pages). The group of peers is presumably going to have high churn (since they go away when you browse somewhere else). Id worry the overhead of keeping track of which peer knows what, especially given how fast the peers change to be a lot. I also expect that for lots of articles, only a very small number of peers will know them.

...

That's true. Dynamically registering / un-registering lookup table gives high overhead on the servers (in both computation & memory usage). Distributed solutions like DHT is there, but we think there could be a trade-off on lookup time for using de-centralized (managing lookup table on the server) versus fully distributed architecture (DHT).

Our prior "naive" implementation costs like if each user has 5K pages cached (with around 50K images), when 10K concurrent user presents it consumes around 35GB of memory, and each registering incurs 500K bytes of network traffic.

We thought it is not that useful, so now we are trying to come up with more lightweight implementation. We hope to have a practically meaningful micro-benchmark result on the new implementation.

Just the metadata for articles, images and revisions is going to be massive. That data itself will need to be distributed too. The network costs associated with just lookups is going to be quite expensive for peers.

It seems your project assumes that bandwidth is unlimited and unrated, which for many parts of the world isn't true.

I don't mean to dissuade you. The idea of a p2p Wikipedia is an interesting project, and at some point in the future if bandwidth is free and unrated everywhere this may be a reasonable way to provide a method of access in case of major disaster of Wikipedia itself. This idea has been brought up numerous times in the past, though, and in general the potential gains are never better than the latency, cost, and complexity associated with it.

• Ryan

---

Wikitech-l mailing list Wikitech-l@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikitech-l

On Mon, Nov 30, 2015 at 4:03 PM, Brian Wolff bawolff@gmail.com wrote:

If we wanted to address such a situation, it sounds like it would be less complex to just setup a varnish box (With access to the HTCP cache clear packets), on that campus.

This is an idea I've casually thought about but never put any real work towards. It would be pretty neat to have something similar to the Netflix Open Connect appliance [0] available for Wikimedia projects.

[0]: https://openconnect.netflix.com/

Bryan

On Sat, Nov 28, 2015 at 1:36 PM, Yeongjin Jang yeongjinjanggrad@gmail.com wrote:

...

*Privacy concerns - Would a malicious person be able to force themselves to be someone's preferred peer, and spy on everything they read, etc.

*DOS concerns - Would a malicious peer or peers be able to prevent an honest user from connecting to the website? (I didn't look in detail at how you select peers and handle peer failure, so I'm not sure if this applies)

Nice points! For privacy, we want to implement k-anonymity scheme on the page access. However, it incurs more bandwidth consumption and potential performance overhead on the system.

Malicious peers can act as if they hold legitimate content (while actually not), or making null request to the peers. We are currently thinking about black-listing such malicious peers, and live-migration of mirror/peer servers if they fails, but more fundamental remedy is required.

Those are interesting ideas, although I'm skeptical you're going to be able to successfully keep malicious peers from tracking users' reading habits, in the same way that law enforcement tracks bittorrent downloads. But it would be great to hear proposals you come up with.

I haven't looked at the code, but are you also preventing malicious peers from modifying the content?